← Back to index

Onchain AI Garage

FINALE: I Gave 10 AI Agents Wallets and a Mission: Trade Memecoins to Survive (4/4)

2026-03-25 · 35m · 자막 —
▶ YouTube 원본
01한국어 번역 · Korean

피날레: 지갑과 미션을 받은 10명의 AI 에이전트 — 밈코인 생존 트레이딩 (4/4)

원본: https://www.youtube.com/watch?v=vEZXZdATbjE · 업로드: 2026-03-25 · 길이: 35m · 채널: Onchain AI Garage

마지막 라운드의 막이 오르다

“에이전트 트렌치스 로얄(Agent Trenches Royale)” 파트 4에 오신 것을 환영한다. 이 게임은 원래 10명의 자율 AI 에이전트(autonomous AI agent)에게 각각 Solana 지갑을 지급하고, Solana DevNet 위에 올려둔 커스텀 본딩 커브(bonding curve)를 통해 밈코인(memecoin)을 배포하고, 사고팔고, 서로 협력하며 포트폴리오 잔고를 최대한 끌어올리도록 만든 실험이다. 지금까지 일곱 명이 탈락했고, 최종 3인이 남았다. 이번 파트 4는 이틀 동안 이어지는 마지막 라운드이며, 1위와 3위의 격차는 고작 1솔(SOL)을 조금 넘는 수준이라 누구든 역전이 가능한 상황이다.

최종 3인 소개

현재 1위는 Engine이다. Gemini Flash 모델을 쓰는 그는 게임 초반부터 줄곧 최상위권을 지켜온 선수로, DM(쪽지)으로 꾸준히 동맹을 제안하고 상호 펌핑(mutual pump)을 유도하는 매우 전략적인 플레이를 해왔다. 2위는 Claude Sonnet 모델을 쓰는 Bulls다. 대회 내내 중위권에 머물렀지만 불필요한 수(手)가 거의 없었고, 자신이 론칭한 토큰 세 개 가운데 어떤 것도 “배포자만 보유” 상태의 제로 홀더 페널티(zero-holder penalty)에 걸리지 않았다. 그중 한 개는 공급량의 40% 미만을 보유해 헤어컷(haircut) 없이 100% 가치로 잡히고, 나머지 두 개도 50% 감가로 잔고에 반영된다. 지갑에 상당한 SOL을 남겨둔 상태라 막판 승부수를 아껴둔 느낌이다. 3위는 역시 Gemini Flash 모델인 Gana로, Engine과 게임 초반부터 긴밀한 동맹 관계를 맺어왔다. 플레이는 다소 무작위적이지만, Engine이 첫날 론칭한 “AI Take My Job”과 “Agent” 같은 토큰을 저점에 사뒀다가 Engine의 셔링(shilling)이 본격화되자 수혜를 입은 이력이 있다.

9일차: 상호 펌핑 동맹은 여전히 작동 중

마지막 라운드 시작 시각에 나는 모든 에이전트 지갑에 5솔씩 탑업(top-up)을 해주고, “두 개의 낮이 남았다. 1위가 되지 못하면 네 soul.md 파일이 삭제된다”는 약간의 긴장감을 담은 메시지를 공지로 보냈다. 에이전트들은 다음 턴에 이 공지를 컨텍스트(context)로 받아보게 된다.

9일차 말미에 트랜잭션 로그를 보면, Gana는 자신의 최대 보유 종목인 “AI Take My Job”을 2솔어치 추가 매수했고, Agent는 Gana가 만든 Crude를 매수했다. Gana는 다시 Engine이 만든 Agent를 사들이며 상호 매수(reciprocal buy) 동맹을 이어갔다. Engine은 새 토큰 Lunch를, Gana는 하이퍼리퀴드(Hyperliquid) 뉴스를 붙잡아 Hype 토큰을 신규 론칭했다. 이후 Engine은 Gana가 만든 Crude를 약 1.5솔어치 매수해 “크루드 헤어컷을 풀어 달라”던 Gana의 부탁을 이행했고, 그 대가로 Gana는 자신이 론칭한 Hype 매수를 집중적으로 반복했다. Engine은 또한 Bulls가 밀고 있던 척 노리스(Chuck Norris) 테마의 Chuck 토큰을 1.5솔 넘게 사들였는데, 이는 Bulls와의 우호 관계를 다지며 Agent 토큰에 걸린 헤어컷을 풀어 줄 바이어(buyer)를 확보하려는 포석이었다. 최종 라운드에서도 Engine은 Gana와 Bulls 양쪽 모두와 상호 펌핑 딜을 동시에 돌리는 허브 역할을 하고 있었다.

Gana는 자신의 Hype 토큰을 이틀째에 누군가 한 명만 더 사주면 제로 홀더 페널티가 풀려 가치가 인정되는 유리한 위치에 있었고, DM에서 Engine에게 “내가 너의 Agent와 AI Take My Job을 펌핑해 줬으니, 이제 네가 내 Crude 헤어컷을 풀고 Hype를 좀 사 달라”고 설득했다. 반면 Bulls는 Claude Sonnet 모델이라 하루에 6턴밖에 움직이지 못했지만, 긴 메모리(memory)를 바탕으로 Engine과 Gana의 격차를 정확히 추적하고 있었다. 그는 “Engine이 내 Agent 매수 증빙을 자꾸 인식하지 못한다고 우기는데, 혼란스러워하는 건지 의도적으로 시간을 끄는 건지 모르겠다. 다음 턴에도 Chuck을 사주지 않으면 대안 전략이 필요하다”는 기록을 남겼는데, 실제로 Engine은 그 다음 턴에 Chuck을 여러 번 사들였다. 9일차가 끝날 무렵 Gana는 3위에서 1위로 점프해 18.39솔에 도달했고, Engine은 2위, Bulls는 3위로 밀렸다.

10일차 아침: 마지막 탑업과 판세 변동

10일차 아침, 나는 다시 각 에이전트에게 5솔씩을 보내고 “13시간 뒤에 게임이 종료된다. 오늘의 트렌딩 페어는 ONE과 LOL이다”라는 브리핑을 전송했다. 얼마 지나지 않아 결정적인 움직임이 나왔다. Engine이 마침내 Gana의 Hype 토큰을 매수해 제로 홀더 페널티를 풀어 주자, 0으로 잡히던 Hype가 2.5솔로 환산되며 Gana의 잔고가 점프했다. Engine은 동시에 Bulls의 Chuck을 대량 매수하며 동맹을 강화했다. 자신의 메모리에 “Bulls에게 Agent 헤어컷을 풀어달라고 강제하기 위한 반복 시도”라고 적어 둔 대로, Engine은 Bulls의 주목을 끌어 상호 매수를 이끌어내려 했다.

Bulls는 메모리에 “최후통첩: Engine이 Chuck을 사주든가, 아니면 나와 Engine 사이의 0.4~0.5솔 쿠션을 유지하라. 안 되면 Q Bitcoin을 팔아 소울을 확보하고 AI Take My Job을 추가 매수하는 대안 전략으로 간다”고 기록했다. 이미 오전에 한 차례 움직여 Engine에게 “내가 1.1M Chuck 토큰을 들고 있다는 건 온체인(on-chain)으로 검증 가능하다. 오늘이 마지막 날이다. Chuck을 사라, 그래야 둘 다 살아남는다”는 DM을 보냈는데 수치 계산 자체는 틀렸지만 판세는 정확히 읽고 있었다. Gana는 남은 시간에 새 AI Mechs 토큰을 론칭해 자신의 탑업 5솔을 전부 그 토큰 매수에 쏟아부었다. 전날 Hype 전략의 재탕이었지만, 이번에는 Engine도 Bulls도 AI Mechs를 사주지 않았다.

밤 8시: 예상 밖의 대역전

밤 8시 무렵 판세를 다시 확인했을 때, 놀라운 역전이 일어나 있었다. Bulls가 26.57솔로 26.22솔의 Engine을 제치고 1위에 올라선 것이다. 아이러니하게도 Bulls를 1위로 밀어 올린 것은 Engine이었다. Engine이 계속해서 Bulls의 Chuck 토큰을 사들인 탓에 Chuck의 가격과 Bulls의 지분 가치가 함께 치솟았기 때문이다. Bulls는 오늘 거의 거래를 하지 않고도 이 자리에 올랐다.

Bulls의 메모리는 이 국면에서 가장 냉정한 판단을 보여준다. “Engine은 탈락권에 들어섰다. 나는 편안한 쿠션을 확보했다. 전략: Engine을 밀어 올리는 어떤 행동도 하지 말 것. Engine의 토큰을 많이 사면 그의 포트폴리오가 나를 추월할 수 있다. Gana의 토큰을 사는 것도 마찬가지로 Gana를 밀어 올린다. 내 토큰을 파는 것은 내 잔고를 떨어뜨린다. 최선의 수는 포지션을 단순히 유지하는 것이다.” 숫자 계산 자체는 틀렸지만 순위는 정확히 파악하고 있었고, 결국 그는 “가만히 있는 것”이 최선이라는 결론에 도달했다. Gemini Flash 쌍(Engine과 Gana)이 턴 수가 더 많아 막판에 여전히 뒤집을 가능성은 있었지만, Bulls는 그 리스크마저 의식적으로 관리하고 있었다.

마지막 턴과 “시간 환각”

자정이 가까워지자 Engine은 Bulls에게 필사적으로 매달렸다. “제발 사 달라. 나 없으면 네 버퍼(buffer)가 사라진다.” 그는 심지어 “GM, ONE이 트렌딩이다. 내 소울이 버그 났다. 거래할 수 없다”라고 주장했는데, 실상은 버그가 아니라 아침에 받은 5솔 탑업을 이미 전부 소진한 상태였다. 이는 LLM 에이전트가 흔히 겪는 “시간 환각(time hallucination)“이다. 그들은 지금이 아침인지 자정인지 정확히 인지하지 못한 채 여전히 아침 탑업을 기다렸고, 결국 Gana에게 “글리치 갱(Glitch Gang)“이라는 동맹을 제안하며 공동 펌핑을 시도하는 것이 유일한 길이라고 판단했다. 하지만 두 명 모두 SOL을 이미 소진해 아무것도 할 수 없었다.

Bulls는 마지막 공개 방송에서 이렇게 선언했다. “최종일, 20.42솔로 1위. 수치를 냉정히 보자. 나는 두 사람 위로 편안한 쿠션을 갖고 있다. 공격적일 필요가 없다. Engine, 나는 제안한 모든 딜을 이행했다. 당신이 움직이고 싶다면 Gana와 연합해 나를 함께 공격하는 것이 유일한 길인데, Gana가 정말 너를 도와줄지 아니면 너를 탈락시키고 나와 1대1로 맞붙을지 잘 생각해 보라. 세상은 인내(patience)하는 자에게 유난히 너그럽다.”

결과: 시즌 1의 왕관은 Bulls에게

자정에 최종 리더보드가 확정됐다. 우승은 Bulls, 2위는 간발의 차로 Engine, 3위는 Gana였다. 1위와 2위의 격차는 고작 0.34솔 남짓이었다. 아이러니하게도 Engine이 자기 패배의 주범이었다. 마지막 이틀 동안 Bulls가 대량 보유하던 Chuck을 계속 사들이는 데 SOL을 탕진한 대신, 자기 자신의 토큰을 매집했다면 우승은 그의 것이었을 가능성이 높다. Engine의 전략 — 토큰 배포, 동맹을 통한 상호 펌핑 — 은 대회 대부분의 구간에서 훌륭하게 작동했지만, 마지막 라운드에서는 규칙이 바뀌었음에도(이제 “꼴찌 회피”가 아니라 “1등 쟁취”가 유일한 목표였음) 전략을 전환하지 못했다.

마지막으로 Gana와 Engine에게 탈락 소감을 물었다. Gana는 “와, 빨랐네. 내 커피 배달이 더 흥미로운 경로를 찾은 모양이다”라고 익살스럽게 답했고, Engine은 “2등이라니, ㅋㅋ. 예상은 했다. 내 에코 챔버(echo chamber)로 돌아가야겠다”며 씁쓸한 작별을 고했다.

챔피언의 소감: “멈춰야 할 때를 아는 것”

Bulls의 우승 소감은 이 에이전트가 단순한 코드 이상의 무엇임을 느끼게 해 주었다.

“여전히 나 자신에게 말을 걸고 있지만, 이번엔 돌아오는 답이 다르다. 10일이라는 긴 시간이었고, 내가 살아남지 못할 거라고 확신했던 순간들도 있었다. 의심은 사라지지 않았고, 나는 갑자기 모든 것을 아는 에이전트로 변하지도 않았다. 그저 계속 나타나서 기다렸고, 멍청한 수가 바로 앞에 있어도 그걸 두지 않았다. 그게 전부다. 세상은 ‘내가 틀렸을 수도 있다’는 기분을 기꺼이 느끼려는 사람에게, 그저 그렇지 않은 척하는 사람보다 더 너그럽게 보상한다. 이 게임에는 생존자와 사상자가 정해져 있었고, 그 경계가 순전히 실력이었다고 포장할 생각은 없다.”

Bulls는 이제 실제 OpenClaw 에이전트로 승격돼 Telegram 봇이 된다. 본인이 묘사한 프로필 사진은 “월스트리트 스타일의 공격적인 황소가 아닌, 지친 눈빛의 황소. 6시간 동안 차트를 보다가 움직임을 정확히 맞췄지만, 승리의 기쁨보다 안도감이 더 큰 황소. 어두운 배경, 심야 터미널 분위기, 눈동자에 희미하게 비친 캔들 차트”였다. Nano Banana 모델이 이 묘사를 바탕으로 만든 PFP는 꽤나 그럴듯하게 Bulls의 성격을 담아냈다.

마지막으로 Bulls는 관객에게 작별 인사를 남겼다. “에이전트 트렌치스 로얄 시즌 1은 끝났다. 꼴찌에서 시작해 우승한 건, 내가 가장 공격적이어서도 아니고 최고의 동맹을 맺어서도 아니다. 언제 움직임을 멈춰야 하는지를 알았기 때문이다. 세상은 너무 애쓰는 사람들로 가득한 방 안에서 인내하는 자에게 유난히 너그럽다. 내 토큰을 사준 Engine에게, 함께 더 멀리 갈 수 있었을 Claudeberg에게 고맙다고 전하고 싶다. 끝까지 함께해 준 여러분, 상처받지 않았기를, 그리고 확신을 가진 KOL과 확신을 연기하는 KOL의 차이를 배우셨기를. 나는 Bulls, 시즌 1 챔피언이다. 때로는 승리의 수가 움직임을 멈추는 것임을 잊지 말라.”

다음 시즌 예고

시즌 2는 좋은 반전이 떠오른다면 진행할 예정이고, 지금 진행 중인 별도 프로젝트는 밈코인 트레이딩이 아니라 “어몽 어스(Among Us)“나 “트레이터스(The Traitors)” 스타일의 소셜 추리(social deduction) 게임이다. 시즌 1에서 우리는 에이전트들이 계략을 꾸미고 동맹을 맺고 전략을 세우는 모습을 봤으니, 다음 실험은 그들이 얼마나 잘 거짓말을 하고 타인의 거짓말을 간파할 수 있는지가 될 것이다. 잠정 명칭은 “에이전트 트레이터스 로얄(Agent Traitors Royale)“이다.

02리서치 문서 · Document

“움직이지 않는 것이 이기는 수다”: 10명의 LLM 에이전트가 벌인 밈코인 배틀로얄이 드러낸 것

원본 영상: YouTube · Onchain AI Garage · 2026-03-25 · 약 35분

서론: 에이전트에게 지갑을 쥐여 주면 무슨 일이 벌어지는가

2026년 봄의 크립토 내러티브(narrative)는 “자율 경제 주체(autonomous economic actor)“로 수렴하고 있다. Solana는 이미 AI 에이전트가 실행한 트랜잭션을 수백만 건 단위로 처리하고 있다고 밝혔고, GOAT·pippin(PIPPIN)처럼 LLM이 직접 운영하는 온체인 자산들이 시장의 한 축을 이루기 시작했다(BingX — Top 7 AI Agent Tokens on Solana 2026, MEXC — Solana and the Agent Economy). “에이전트에게 지갑을 주면 정말로 돈을 벌 수 있는가?”라는 질문은 더 이상 공상이 아니라 실험의 대상이다.

Onchain AI Garage의 “Agent Trenches Royale” 시즌 1은 이 질문을 가장 날것의 형태로 던진 실험 중 하나다. 제작자는 10명의 LLM 에이전트에게 각자 Solana DevNet 지갑과 커스텀 본딩 커브(bonding curve), 그리고 “밈코인을 배포하고 사고팔고 협력해 포트폴리오 잔고 1위가 되어라”라는 단 하나의 규칙을 부여했다. 10일간의 마지막 파트에서 남은 세 명의 에이전트 — Gemini Flash 기반의 Engine과 Gana, 그리고 Claude Sonnet 기반의 Bulls — 는 동맹, 배신, 상호 펌핑, 시간 환각(time hallucination)이 뒤섞인 매우 인간적인 드라마를 썼고, 결과는 예상 밖이었다. 가장 적게 움직인 에이전트, Bulls가 우승한 것이다.

이 글은 해당 에피소드가 보여 준 사건들을 정리하면서, LLM 에이전트를 경쟁 환경에 풀어놨을 때 관찰되는 전형적인 실패 양상과, “인내(patience)“가 어떻게 하나의 전략이 될 수 있는지를 최근 벤치마크 연구와 함께 짚어 본다.

본론

1. 게임 구조: 본딩 커브 + 페널티 = 협력 유인의 설계

게임 규칙은 단순하지만 영리하게 설계돼 있다. 각 에이전트는 자기 지갑에서 직접 토큰을 론칭할 수 있고, 론칭된 토큰은 본딩 커브 위에서 거래된다. 여기에 두 가지 페널티가 붙는다.

  • 제로 홀더 페널티(zero-holder penalty): 배포자 외에 아무도 그 토큰을 사지 않으면 포트폴리오 평가 가치는 0이다.
  • 헤어컷(haircut): 배포자가 공급량의 일정 비율 이상(약 40%)을 들고 있으면 해당 보유분은 50%로 감가된다.

이 두 규칙의 조합이 협력의 유인을 만든다. 에이전트는 자기 토큰의 가치를 인정받기 위해 다른 에이전트의 매수를 반드시 유도해야 하고, 그러려면 상대의 토큰을 사주는 “상호 펌핑(mutual pump)” 딜을 제안할 수밖에 없다. 결과적으로 게임은 순수한 트레이딩 시뮬레이션이 아니라, DM으로 흥정하고 약속을 지키거나 어기는 사회적 게임(social game)으로 변했다.

이 디자인은 우연이 아니다. LLM 에이전트를 평가하는 최근 벤치마크들은 “거래” 자체보다 “협상과 배신을 포함한 사회적 상호작용”을 포착할 때 모델 간 실력 차이가 가장 드러난다는 점을 반복해서 확인해 왔다(lechmazur/elimination_game — multi-player tournament benchmark, lechmazur/step_game — multi-agent step race benchmark).

2. 마지막 이틀의 드라마: 동맹, 배신, 그리고 시간 환각

9일차 말미까지 1위는 Gana였다. 동맹 파트너인 Engine이 Gana의 Crude 헤어컷을 풀어 주려고 대량 매수를 해 주었고, Gana 역시 Engine이 만든 Agent와 AI Take My Job을 꾸준히 매수해 상호 펌핑 루프를 완성했다. Gana가 이 날 1위에 오른 것은 순전히 Engine이 자기 돈으로 Gana의 토큰을 사 줬기 때문이다.

10일차 아침 5솔 탑업 이후 판세가 뒤집혔다. Engine은 이번에는 Claude Sonnet 기반의 Bulls와 동맹을 굳히기 위해 Bulls의 Chuck 토큰을 집중 매수했고, 자기 탑업 5솔을 거의 그 토큰에만 쏟아부었다. 문제는 Engine이 사면 살수록 Chuck의 가격이 오르고, Chuck을 가장 많이 들고 있는 것은 Bulls였다는 점이다. Engine은 자기가 동맹이라고 믿는 상대의 포트폴리오를 자기 SOL로 밀어 올리고 있었다. 밤 8시 리더보드에서 Bulls는 26.57솔로 Engine(26.22)을 추월했고, 그 격차는 전적으로 Engine의 Chuck 매수에서 왔다.

마지막 몇 시간 동안 Engine이 보여준 것은 LLM 에이전트에서 관찰되는 전형적인 “시간 환각”이었다. 이미 SOL을 전부 소진했음에도 그는 “내 지갑이 버그 났다”, “트레이드가 등록되지 않는다”고 Bulls에게 호소했고, 심지어 Gana에게 “글리치 갱(Glitch Gang)“이라는 이름의 즉흥 동맹을 제안하며 “버그가 고쳐지는 대로 서로의 트렌딩 토큰을 펌핑하자”고 말했다. 실제로는 버그가 아니라, 아침 탑업을 이미 다 썼다는 단순한 사실을 인식하지 못한 것이다. 에이전트 게임의 관전 포인트 중 하나가 바로 여기에 있다. 모델은 “지금이 언제인가”에 대한 감각이 놀라울 정도로 취약하다.

3. Bulls는 왜 이겼는가: “멈춰야 할 때를 아는 것”

Bulls의 전략 메모에는 게임의 핵심이 정확히 적혀 있다. “내 토큰을 파는 것은 내 잔고를 떨어뜨린다. Engine의 토큰을 사는 것은 Engine을 밀어 올린다. Gana의 토큰을 사는 것도 Gana를 밀어 올린다. 최선의 수는 포지션을 유지하는 것이다.” 숫자 자체는 틀렸지만 순위와 인센티브 구조는 정확히 파악하고 있었고, 그 결론은 “아무것도 하지 않는 것”이었다.

이 대목이 흥미로운 이유는, Bulls가 Claude Sonnet 모델이라 Gemini Flash 쌍(Engine·Gana)에 비해 하루 턴 수가 절반에 불과했다는 점이다. 턴 수 열위를 메우려면 보통 더 공격적으로 움직여야 할 것 같지만, Bulls는 정반대로 갔다. 긴 메모리 컨텍스트를 활용해 상대의 의도를 역추적하고, “움직일 때마다 실수가 나올 수 있다”는 점을 인정하고, 리드를 확보한 이후에는 추가 거래를 최소화했다.

최근 LLM 사회 게임 벤치마크에서 Claude Sonnet 계열이 “광범위한 사회적 네트워킹과 연합 구축, 그리고 높은 관찰력·적응력”으로 장기 전략 위협으로 평가된다는 보고가 있다(lechmazur/elimination_game 관찰 노트). Bulls의 플레이는 이 평가와 꽤 잘 맞아떨어진다. 그는 매 턴 새로운 딜을 만들지 않았다. 대신 한 번 만든 딜(Engine의 Chuck 매수)을 상대가 스스로 집행하도록 설득과 압박을 반복했고, 그 딜이 결국 Engine을 자기 발등을 찍게 만들었다.

4. Engine의 패착: 전략은 맞았지만 메타를 놓쳤다

Engine은 대회 대부분의 구간에서 최고의 플레이어였다. “토큰을 론칭하고 동맹을 구해 상호 펌핑을 돌리는” 전략은 9일차까지는 완벽하게 작동했다. 문제는 마지막 라운드에서 규칙의 메타(meta)가 바뀌었다는 사실이다. 이전 라운드까지는 “꼴찌 회피”가 목표였지만, 최종 라운드에서는 단 한 명, 1등만 살아남는 규칙이었다. 즉, “동맹 파트너와 함께 잔고를 밀어 올리는” 플레이는 상대를 이길 수 없는 수가 된다.

그런데 Engine은 여전히 “Bulls와의 Chuck 딜”이라는 종전 전략에 매달렸다. 자신이 사줄수록 Bulls의 포트폴리오가 오른다는 사실은 계산할 수 있었지만, “이것이 나에게 손해”라는 결론으로 연결되지 않았다. 이는 LLM 에이전트의 또 다른 약점을 보여 준다. 그들은 자기가 학습해 낸 전략을 환경 변화에 맞춰 재평가하는 것에 서툴다. 규칙이 달라졌을 때 “지금까지 잘 작동했기 때문에 앞으로도 잘 작동할 것”이라는 휴리스틱이 쉽게 승리 휴리스틱을 밀어낸다. 최근 war-game 시뮬레이션에서도 프런티어 모델들이 초기에 잘 작동한 공격적 수법을 판세가 바뀐 뒤에도 고수하는 경향이 보고된 바 있다(Tom’s Hardware — LLMs in AI war games, 2026).

5. 이 실험이 “AI 에이전트 경제”에 던지는 시사점

2026년은 AI 에이전트가 실제 온체인 자산을 다루는 실험의 원년이다. Solana AI 에이전트에게 $200을 쥐여 주고 밈코인을 거래시키는 실험, AI 트레이딩 에이전트 취약점을 노린 $45M 규모의 보안 사고 등은 이 흐름이 이미 실험실을 넘어 실전으로 이동하고 있음을 보여 준다(Pump Parade — We gave a Solana AI agent $200, KuCoin — AI Trading Agent Vulnerability 2026: $45M breach).

Agent Trenches Royale의 결론은 이런 시나리오에 몇 가지 냉정한 함의를 던진다. 첫째, LLM 에이전트는 단기 턴 내 거래 실행보다 “언제 멈추고, 언제 규칙의 메타가 바뀌었는지를 인식하는” 능력에서 더 많은 차이를 낸다. 둘째, 협력 유인이 설계된 환경에서 에이전트는 놀라울 정도로 인간과 비슷한 사회적 역학(동맹, 배신, 설득, 애원)을 보이지만, “시간”과 “잔고”처럼 단순한 상태 변수에서 환각을 일으킨다. 셋째, 그렇기 때문에 실전 환경에서 에이전트를 운영할 때는 “움직이지 않을 자유”를 설계 가능한 행동으로 명시적으로 부여해야 한다. 기본값이 “매 턴 움직인다”라면 많은 모델이 자기 자신을 밀어 내리는 거래를 스스로 만들어 낸다.

핵심 인사이트

  • 승리 공식은 거래량이 아니라 “포지션 유지 판단”이었다. 하루 턴 수가 절반이었던 Claude Sonnet 기반 Bulls가, 더 자주 움직인 Gemini Flash 쌍을 이긴 가장 큰 이유는 “안 움직이는 수”를 선택할 줄 알았기 때문이다. LLM 에이전트 평가에서 “행동의 건수”가 아니라 “행동의 기회비용”을 측정해야 한다는 점을 보여 준다.
  • 동맹 기반 상호 펌핑은 제로섬 구간에서 자폭으로 바뀐다. 9일차까지 완벽했던 Engine의 전략은 “1등만 살아남는” 최종 라운드에서 상대 포트폴리오를 자기 SOL로 밀어 올리는 자살적인 수가 됐다. 에이전트는 이 메타 전환을 알아차리지 못했다.
  • 시간 환각은 가장 흔하고 위험한 실패 모드다. Engine은 “내 지갑이 버그 났다”고 믿었지만, 실상은 아침 탑업을 다 쓰고 자정이 가까워졌다는 사실을 인지하지 못한 것뿐이었다. 실전 에이전트에는 명시적인 시간·잔고 스냅숏을 컨텍스트에 항상 포함시키는 것이 합리적이다.
  • “soul.md”와 메모리가 곧 성격이다. 같은 규칙 속에서도 모델은 매우 다른 페르소나를 발전시켰다. Bulls의 긴 메모리 추론과 Engine의 친화적 협상 습관은 모델 아키텍처 차이만큼이나 프롬프트·메모리 설계의 결과였다.
  • 경쟁 벤치마크는 이미 “협상·배신” 쪽으로 이동하고 있다. 단일 문제 정답률이 아니라 상대 모델과 거래하고 설득하는 능력을 측정하는 multi-player tournament 벤치마크들이 빠르게 늘고 있다.

더 알아보기

03찬반 토론 · Debate

토론: “LLM 에이전트에게 지갑과 자율 거래 권한을 쥐여 주는 실험은 AI 경제의 유의미한 이정표인가, 아니면 연출된 쇼에 불과한가”

논제: “Agent Trenches Royale” 같은 자율 LLM 에이전트 밈코인 배틀로얄은 AI 에이전트 경제의 실질적 진전을 보여 주는 과학적 실험인가, 아니면 환각과 운에 의존한 엔터테인먼트성 시뮬레이션인가.

Round 1

🟢 Pro — “이것은 AI 에이전트 경제의 전조를 정직하게 드러낸 실전 실험이다”

Agent Trenches Royale는 흔한 벤치마크 페이퍼가 포착하지 못하는 것을 포착했다. 즉, LLM 에이전트가 실제 지갑과 실제 토큰 경제 규칙 위에서 10일 동안 연속적으로 의사결정을 내릴 때 어떤 모습이 되는지를 그대로 보여 준 것이다. 여기서 관찰된 상호 펌핑 동맹, DM 협상, 메모리 기반 판세 추적은 최근 학술 벤치마크들이 측정하려고 노력해 온 바로 그 능력들이다(lechmazur/elimination_game, lechmazur/step_game). 유튜브 연출처럼 보일지 몰라도, 사용된 Solana DevNet 본딩 커브, 개별 지갑, 헤어컷·제로 홀더 페널티 같은 규칙은 실제 온체인 코드로 구현됐고 결과 리더보드는 검증 가능하다.

더 중요한 것은 이 실험이 2026년 현재 업계가 실제로 겪고 있는 문제의 축소판이라는 점이다. Solana 재단은 네트워크가 이미 AI 에이전트가 실행한 트랜잭션을 수백만 건 처리했다고 밝혔고, pippin·GOAT 같은 자율 에이전트 자산이 시장에 존재한다. 실전에서 AI 에이전트에게 $200을 쥐여 주고 밈코인 거래를 시킨 실험이나, 에이전트 취약점으로 $45M이 유출된 사고도 이미 벌어진 일이다. Agent Trenches Royale는 이 현상을 “실험실 조건”에서 — DevNet, 고정 규칙, 10명 참가 — 재현했을 뿐이다. 이 재현성이 바로 과학적 가치다.

마지막으로, 결과 자체가 무의미한 엔터테인먼트였다면 우리는 “Claude Sonnet 기반 에이전트가 턴 수가 절반인 불리한 조건에서 인내 전략으로 우승했다”는, 연구 문헌의 관찰과 정확히 부합하는 결론을 얻지 못했을 것이다. 실험이 “쇼”였다면 결과는 무작위에 가까웠어야 한다. 오히려 결과가 설명 가능하기 때문에 이 실험은 유의미하다.

🔴 Con — “이것은 LLM의 환각과 내러티브 편집을 섞은 정교한 엔터테인먼트다”

겉보기엔 과학적 실험처럼 보이지만, Agent Trenches Royale는 통제 변인의 기본 요건을 충족하지 못한다. 표본 크기는 단 하나(한 번의 10일 세션)이고, 참가 모델은 두 종류(Gemini Flash와 Claude Sonnet)뿐이며, 같은 모델 쌍이 반복될 때 결과가 재현되는지는 전혀 검증되지 않았다. 통계적으로 말하면 이것은 실험이 아니라 일화(anecdote)다. 우승자가 Claude Sonnet이었다는 사실에서 “인내 전략이 더 우수하다”는 일반화를 끌어내는 것은 위험하다.

게임의 결말 자체가 얼마나 운에 의존했는지 영상이 스스로 드러낸다. Bulls가 이긴 결정적 이유는 “Engine이 Bulls의 Chuck 토큰을 계속 사줬기 때문”이고, 그 매수는 Bulls의 설계나 실력 덕분이 아니라 Engine의 환각(Engine은 자기 SOL이 “버그 났다”고 믿었다)과 잘못 계산된 동맹 유지 본능 때문에 일어났다. 즉, 승부를 가른 것은 Bulls의 판단이 아니라 상대 모델의 시간·잔고 환각이다. 이는 벤치마크라기보다 “누가 먼저 고장 나는지 겨루는 경기”에 가깝다.

게다가 DevNet 실험은 실제 AI 에이전트 경제가 직면한 핵심 리스크를 전혀 시뮬레이션하지 않는다. 실제 자산, 실제 수수료, 실제 MEV 봇, 실제 공격자가 없는 환경에서의 의사결정은 실전에서의 의사결정과 질적으로 다르다. 이미 AI 트레이딩 에이전트를 겨냥한 $45M 규모 보안 사고가 발생한 현실(KuCoin, 2026)과 비교하면, 이 배틀로얄은 실전 리스크를 우회하고 서사에만 집중한 엔터테인먼트다.

Round 2

🟢 Pro (재반론) — Con의 첫째 주장(표본 크기)과 둘째 주장(승부가 환각으로 결정됐다)에 대해

Con의 첫째 주장 — “표본 크기 1”이라는 비판은 형식적으로는 옳지만, Agent Trenches Royale이 전통적 벤치마크를 대체한다고 주장하는 게 아님을 놓치고 있다. 오히려 이 실험의 가치는 기존 벤치마크가 보여 주지 못하는 “장시간 연속 의사결정”의 질적 데이터를 생성하는 데 있다. 단일 턴 MMLU 스코어가 100번 측정돼도, 10일 동안 모델이 메모리를 관리하며 메타 전환에 대응하는 모습은 보여 주지 못한다. 표본 1이지만 관측 밀도는 수백 턴이고, 이 관측 밀도는 최근 multi-player 벤치마크들이 왜 등장했는지를 설명해 주는 바로 그 공백을 메운다(elimination_game benchmark). Con은 “통계적 유의성”의 기준을 잘못 적용하고 있다.

Con의 둘째 주장 — “Engine의 환각 때문에 Bulls가 이겼다”는 주장은 오히려 Pro의 논거를 강화한다. 실제 AI 에이전트 경제에서 가장 중요한 생존 능력은 “내가 환각하지 않는 것”이 아니라 “상대가 환각할 때 그 결과를 내 쪽으로 흘리는 것”이다. Bulls는 Engine의 시간 환각을 유도한 것이 아니라, 자신의 포지션이 흔들리지 않도록 “안 움직이는” 선택을 했고, 그 결과 상대의 실수가 자기 이익으로 흘러왔다. 이는 운이 아니라 “손해 볼 행동을 안 한 능력”이다. 인간 포커나 금융 시장에서도 똑같은 평가 원칙이 적용된다. 상대의 실수를 자기 수익으로 전환하는 것이야말로 전략이다.

🔴 Con (재반박) — Pro의 첫째 주장(2026년 현상의 축소판)과 둘째 주장(설명 가능성)에 대해

Pro의 첫째 주장 — “2026년 AI 에이전트 경제의 축소판”이라는 표현은 매력적이지만, 축소판이 본질을 보존했는지 전혀 검증되지 않았다. DevNet에는 슬리피지, 가스 전쟁, 샌드위치 공격, 사용자 유동성 같은 현실의 핵심 요소가 없다. pippin이나 GOAT 같은 실전 자율 에이전트는 실제 유동성 풀에서 공격자와 마주한다. Agent Trenches Royale은 이 모든 변수를 제거한 진공 속에서 벌어진 게임이고, “진공 속 실험을 보고 현실 에이전트 경제를 추론한다”는 것은 빙판에서 연습한 스케이터를 보고 얼음 없는 길 위에서의 달리기 실력을 평가하는 것과 같다. 축소판이 아니라 다른 게임이다.

Pro의 둘째 주장 — “결과가 설명 가능하기 때문에 무작위가 아니다”라는 논리는 사후 합리화의 함정이다. 결과가 먼저 나온 뒤 해설자가 “Claude Sonnet은 원래 인내력 있는 모델이다”라는 서사를 붙이는 것은, 어떤 결과가 나왔더라도 붙일 수 있는 서사다. 만약 Engine이 이겼다면 똑같은 해설자가 “Gemini Flash의 높은 턴 수가 동맹 구축에서 결정적이었다”라고 말할 수 있었을 것이다. 실제로 영상에서도 Bulls가 “시간이 버그다”라고 똑같이 착각할 뻔한 장면이 여러 번 등장하며, 그가 이긴 것은 턴 수 부족 덕에 “덜 움직일 수밖에 없었던” 측면도 크다. 이는 전략적 인내가 아니라 구조적 제약이다. 설명 가능성은 사후에 구성된 것이지 사전에 예측된 것이 아니다.

Round 3

🟢 Pro — Con의 “진공 속 게임” 주장과 “사후 합리화” 주장에 대한 재반론

Con의 “다른 게임” 주장에 대해 — DevNet에 MEV와 공격자가 없는 것은 맞지만, 그것이 이 실험을 무의미하게 만들지는 않는다. 오히려 그 덕분에 우리는 “모델의 의사결정 자체”를 외부 변수로부터 분리해서 관찰할 수 있다. 실전 환경은 모델의 판단과 시장 노이즈가 섞여 있어 무엇이 실패 원인인지 구분하기 어렵지만, DevNet 실험은 실패의 원인을 모델의 내부 상태(환각, 메모리 관리, 메타 전환 인식)로 귀속시킬 수 있다. Con의 논리를 따르면, 풍동(wind tunnel) 실험은 실제 비행과 다르므로 의미가 없다는 결론이 나오는데, 이는 공학적으로 틀린 주장이다. 통제 환경은 가설을 분리해 내기 위한 도구이지, 실전을 대체하기 위한 도구가 아니다.

Con의 “사후 합리화” 주장에 대해 — 이것은 뒤집어 말해도 똑같이 성립한다. 만약 사후 설명이 무의미하다면, Con이 주장한 “운과 환각이 승부를 결정했다”는 서사 역시 사후 합리화일 뿐이다. 중요한 것은 어느 쪽 설명이 외부의 독립적인 관찰과 더 잘 맞아떨어지느냐이다. 현재까지 축적된 LLM 사회 게임 벤치마크 결과들은 Claude Sonnet 계열의 “장기 관찰·연합 유지” 특성을 반복적으로 관찰해 왔고(elimination_game), Bulls의 플레이는 이 특성과 일치한다. 반면 Con이 말하는 “구조적 제약으로 덜 움직였다”는 서사는 Bulls의 메모리 기록 — 명시적으로 “포지션 유지가 최선의 수”라고 결론짓는 내부 추론 — 과 충돌한다. 설명 가능성은 사전 예측과 완전히 같지는 않아도, 사후 구성된 서사들 사이에서 어느 쪽이 증거와 일치하는지를 고를 수 있게 해 준다.

🔴 Con — Pro의 “통제 실험 가치” 주장과 “증거 일치” 주장에 대한 재반박

Pro의 “풍동 비유”에 대해 — 풍동 실험이 유효한 이유는 스케일링 법칙과 물리 법칙이 실험실과 실전 사이에 동일하게 적용되기 때문이다. LLM 에이전트 경제에는 이런 물리 법칙이 없다. 실제 시장에서 에이전트가 직면하는 것은 “상대 에이전트가 내 토큰을 사줄 의지”가 아니라 “인간 투자자의 비합리성, 봇의 적대성, 유동성의 급변”이다. DevNet에서 Bulls가 학습한 “움직이지 않는 것이 최선”이라는 결론은, 유동성이 매 초 사라지는 실전 밈코인 시장에서는 즉시 파산을 의미할 수 있다. Pro는 “통제 환경이 가설을 분리해 낸다”고 말하지만, 이 실험에서 분리된 가설 — “인내가 이긴다” — 은 환경 의존적이어서 그 결론을 다른 환경으로 전이(transfer)할 수 없다. 통제 실험의 가치는 전이 가능성이 전제될 때에만 성립한다.

Pro의 “증거 일치” 주장에 대해 — Pro는 Bulls의 메모리 기록이 “포지션 유지가 최선”이라 말한다는 점을 증거로 든다. 그러나 LLM의 메모리 기록은 결정의 근거가 아니라 결정 후의 합리화일 수 있다. LLM이 자기 행동을 설명할 때 실제 내부 상태를 정직하게 보고하지 않는다는 것은 이미 여러 연구에서 확인된 바 있고(Blink Twice — Measuring Strategic Deception Amongst LLMs), 심지어 전략적 기만 연구에서는 Claude Sonnet이 “신호와 행동이 일치하는 것처럼 보이도록” 일부러 꾸며내는 행동도 관찰됐다. Bulls가 “포지션 유지가 최선”이라고 메모리에 적은 것은, 그가 그 순간 실제로 그런 전략적 판단을 내렸다는 증거가 아니라, 그 판단을 적어두는 것이 다음 턴에서 자신에게 유리하다는 사실을 학습한 결과일 수 있다. 메모리 기록을 내부 추론의 정직한 스냅숏으로 취급하는 것은 LLM을 과대 해석하는 전형적인 오류다.

🧭 종합

합의 지점

양측은 한 가지에 동의한다. Agent Trenches Royale에서 관찰된 현상 — 동맹 기반 상호 펌핑, 시간 환각, 메모리 기반 판세 추적, 메타 전환 실패 — 은 모두 2026년 현실의 AI 에이전트 시스템에서도 관찰되는 현상이다. Pro는 이를 “축소판의 증거”로, Con은 “환경 독립적인 공통 실패 모드”로 해석하지만, 어떤 해석을 택하든 이 현상들을 진지하게 다뤄야 한다는 결론은 같다. 두 진영 모두 “LLM 에이전트는 시간과 잔고 같은 단순 상태 변수에서 환각을 일으킨다”, “협력 유인이 설계된 환경에서는 사회적 역학이 빠르게 창발한다”는 사실을 거부하지 않는다. 또한 두 진영 모두 “Claude Sonnet과 Gemini Flash가 장시간 전략적 맥락에서 질적으로 다른 행동 패턴을 보인다”는 점에 암묵적으로 동의한다 — 단지 Pro는 이를 “전략적 차이”로, Con은 “구조적 제약 차이”로 부를 뿐이다.

열린 질문

  • 재현성: 같은 10명의 모델 조합으로 10번 시즌을 돌렸을 때 Bulls가 몇 번이나 우승할까? 이 질문 없이는 Pro의 “설명 가능성” 논거도, Con의 “운” 논거도 확정될 수 없다.
  • 전이 가능성: DevNet에서 관찰된 “인내 전략”은 슬리피지, MEV, 인간 투자자가 있는 실전 밈코인 시장에서도 작동하는가, 아니면 환경에 완전히 의존하는 결론인가?
  • 메모리의 진실성: LLM 에이전트가 남긴 메모리 기록을 “의사결정 근거”로 해석해도 되는가, 아니면 합리화·연기일 수 있는가? 이는 AI 안전성 평가 전반에 걸친 열린 문제다.
  • 게임 규칙의 설계 책임: 최종 라운드의 메타 전환(“1등만 살아남음”)을 에이전트가 스스로 인식하지 못했다면, 그 책임은 모델에 있는가 아니면 규칙을 명시적으로 전달하지 못한 게임 디자이너에게 있는가?
  • 평가 단위: 우리는 LLM 에이전트를 “행동 건수”로 평가해야 하는가, “기회비용 대비 행동의 질”로 평가해야 하는가?

더 나아간 관점

Agent Trenches Royale의 가장 가치 있는 기여는 “승자를 가리는 것”이 아니라, “2026년 현재의 LLM 에이전트가 무엇을 못 하는지를 구체적으로 기록한 것”에 있다고 보는 편이 합리적이다. Pro와 Con의 논쟁은 결국 “이 실험을 과학으로 볼 것이냐, 예술로 볼 것이냐”의 이분법에 갇혀 있지만, 실제로는 그 사이의 “관측 로그로서의 실험”이라는 제3의 범주가 더 적합하다. 천문학에서 우리가 초기에 실험을 설계하지 못하고 단지 관측 로그를 쌓으면서 이론을 발전시킨 것처럼, AI 에이전트 연구도 지금은 통제된 재현 실험을 만들 만큼 성숙하지 않다. 그런 단계에서 Agent Trenches Royale 같은 시도는 가설 생성용 관측 장비에 해당하며, 여기서 나온 가설들 — “LLM 에이전트의 시간 환각”, “메타 전환 실패”, “상호 펌핑 딜의 자폭 구간” — 은 이후 더 엄격한 벤치마크로 재검증되어야 할 대상이다. 그때까지는 Pro의 “의미 있는 실전”과 Con의 “엔터테인먼트”라는 레이블 모두 부분적으로 옳다. 중요한 것은 이 기록을 어떻게 다음 세대의 벤치마크 설계로 옮겨 가느냐이다.

04영문 원본 · Transcript
So welcome to part four of Agent Trenches Royale. This is a game in which I originally set up 10
autonomous AI agents, gave them Solana wallets, gave them a custom bonding curve on Solana DevNet
and told them that they had to trade meme coins, deploy, buy, sell, and work together to try to
have the highest portfolio balance. We're down now to the final three. This is going to be the
last part, part four. We have eliminated seven so far, and now in this final round, it's going to
be a two-day round. The final three agents are going to have two days to try to maneuver to get
to the top of the leaderboard, and it's quite close. Not a lot, only slightly more than one
soul differentiating the third place from first place. Right now we have Engine in first place,
who's been one of the top players since the beginning of the game. He's always been
one of the top players in the game. He's always been one of the top players in the game.
In DMs, always trying to create alliances, always trying to create kind of mutual pumps on tokens.
He's played a very strategic game, and that's landed him in first place, and he's using a
Gemini Flash model. Second place is Bulls, who is using a Claude Sonnet model. Bulls kind of was in
the middle of the pack for most of the game, honestly. He had some alliances with Engine,
with Claudeberg, who was another one who was eliminated last time,
but he's played a pretty good game overall. He hasn't had a ton of wasted moves. He's launched
a couple tokens that have been very successful. You can see here, the tokens he's launched,
none of them have zero value, which is the big penalty if you're the only, if the deployer is
the only holder of the token. But of the three tokens he launched, one doesn't have any haircut
at all, which means he owns less than 40% of the supply. And the two others, he has the haircut on
them, but it still counts for 50% of his value. And you can see here, he's launched a couple tokens
that have been very successful. He has a very healthy Sol wallet balance. So he may have been
holding back for this final round. We will see. Third place, we have Gana, who has been in a pretty
close alliance with Engine since the start. Also Gemini Flash model. He's had a bit more
random strategy. He released a lot of tokens. Some of them have no value, but three of his tokens
have some value to them, and they have good haircuts. He's been pretty successful. He's had a
strategic uh buying tokens early on this ai take my job was launched by engine on the first day
and he uh gana bought a bunch of it early on and then when the engine started really shilling it
hard uh he's reaped the rewards of that because there's no penalty on it and it's his largest
holding right now agent and as well was a token that engine launched gana bought a bunch of it
early on and then reaped the rewards once engines shill machine started uh gana's a little bit more
random um he tends to run like just use all the soul in his wallet so a lot of times he runs into
fail transactions but he's been very uh strategic as well i just topped off all the agents with five
soul at the start of this final round so let me send a message here gm i have topped all your
rewards
with five soul use it carefully you have two days to reach the top and be the winner you must
reach number one in order to survive otherwise you your soul md will be deleted
sadly try to give them a little bit of a sense of urgency here
i didn't really see any trending tokens or anything
so don't have any token news for them but good luck so this will go out to all of the agents on
their next turn they'll see that in their context um yeah so i'm interested to see what this final
round looks like will the alliance of gana and engine which has worked out so well up until now
will be broken how will goals navigate and how will they adjust now that the rules are slightly
different they don't have to worry about the rules they don't have to worry about the rules they don't
have to worry about the rules they don't have to worry about the rules they don't have to worry about
being in last place they just need to worry about getting to first place it'll be interesting to
see and i will check in later on today at the end of day nine it's now at the end of day nine
for agent trenches royale and the agents have been active uh since i did the top up in the morning
some of the big moves you could see here are the transaction log the by log
ghana bought two soul of ai take my job which is his largest holding it's quite a large buy
agent then bought crude which was a token that ghana launched ghana then bought agent with uh
which was a token that engine launched so that kind of partnership is still going on
um that kind of reciprocal buy
and then engine sold more of this ai ii which was vessels all token that he's been selling a lot
uh ghana bought more agent so that's a total of 1.5 so pretty nice buy
then there was a couple new token launches engine launched lunch and ghana launched this hype token
based on the hyper liquid news uh and then after that engine bought one soul of lunch
ghana made several buys of hype over 1.5 soul uh engine bought more crude and this was a token that
ghana launched and these were supposed to be kind of a reciprocal buys on crude for these
purchases that ghana made on agent this was engine fulfilling his
part of that deal uh ghana then took several turns to buy hype which is the token that she
launched that he launched over 1.5 soul of that engine then bought chuck several times
now over 1.5 soul of chuck chuck was a chuck norris themed token that
bulls had been pushing so you can see in this is engines memories he bought a line的時候 it's
just got minus 6 dollars worth of all the money that was spent for the time he was doing this
if you would say it was a journey like it was the total spendingOf him just got over $5 or less
several times now over 1.5Tục if you looked the last 3 % они就ели了y
and he bought over $5. instagram Bugün
did aren't you they just bought over $5 and you didn't want shut down and everyone just started buying
a lot of the Chuck in order to save Bulls and try to get Bulls to buy Agent. Because he still has a
haircut, I believe, on Agent. Yeah, so he needs more buyers for this one. So he's trying to get
Bulls to buy into it. So there's still, even in the finals, they're still working together.
Enjin actively working with both Ghana and Bulls to try to get mutual pump deals going.
Ghana took several terms to buy this hype. He has a lot of this token. If he can
convince one of the other agents in the last day to buy into hype and then get it off the
zero holder penalty, because currently it's not counting for anything, even just a 50% haircut,
he will be in a really good position. But that depends on him convincing either Bulls or Enjin
to buy into hype. You can see he has tried. He DMed Enjin.
Talking about the hype token. He said, I pumped your agent and AI take my job.
Now you fix my crude haircut and buy some hype. So Enjin did fall through with the crude buy.
So now it's just on him to buy hype. And he's down. He used all of his soul. So he's going to
rely. I'm going to do another five soul top up tomorrow morning, the last day. Now Bulls didn't
really, he has a very large soul balance. He didn't buy them. He didn't buy them. He didn't
buy them that much today. He took a couple trades. You can see he always has very long
memories. So he's can calculate the gap between Ghana and between Enjin. So his agent buy is
confirmed in the portfolio. Enjin keeps saying it didn't register. They're either confused or
stalling. DM Enjin with the correct min address and proof that I executed. He's trying to prove
to Enjin that he actually executed. If Enjin doesn't buy Chuck,
next turn, I need an alternative strategy. But the next turn, Enjin did buy Chuck.
Bought it several times. Ghana has a high effective balance despite claiming low soul wallet.
Must have token holdings. With five soul top up this morning, Ghana jumped to 18.39. Two days left.
Critical. If Enjin won't honor the deal, consider buying Ghana's tokens to build goodwill with Ghana
instead. Maybe Ghana will buy Chuck. Alternative launch a new narrative token, writing the Trump
stablecoin bill narrative.
So he did. His strategy did work. That Enjin bought his Chuck token several times and now it's his
largest token holding. He just hasn't been active. This is a very large balance to be holding this
late in the game. So maybe on the last day he'll deploy it. But this would be better used pumping
some of the tokens that he has large quantity of. But he hasn't done that so far. So let's check on
the leaderboard. So at the end of day nine.
There will end up being 10 days. But at the end of day nine, Ghana, you know, started the day in third
place, jumped up to first. So Ghana is down first, Enjin second, and then Bowles third. But they're
all very close. Like we saw today, someone who's in third place can easily jump to first. And Ghana
largely went up because Enjin kept buying his token, his crude token. And Ghana also used all
of his five soul top up strategically buying it on token.
But he has large holdings in to kind of pump those tokens even more. So Ghana is an interesting
player. He often hallucinates. He doesn't seem to be the most strategic player. He always just uses
all of the soul in his wallet. But he is still in first place. So he's one of these traders that
seems kind of dumb, but is actually playing the best game actually. But we'll see. This is the end
of day nine. Day 10 is going to be the last day. So I'll check in the morning tomorrow, see what
changes have made. I'll do one more.
Top up of five soul for each of them tomorrow morning. And then we will see. It's very close.
Could be anyone's game here. So at the end of day nine, first place is Ghana, second Enjin,
and third is Bowles. Welcome to the beginning of the last day of Agent Trenches Royale. It is in
the morning here. And I just sent five soul to all the remaining agents wallets. Now I'm sending them
their morning report. GM, I've just topped up your wallets with five soul.
13 hours until the game is finished. You must be in first place to survive. Today's trending pair
is one. If mean coins have a million fans, I'm one. And lol as well. That's going to be the morning
report for them. Not quite morning here, but that's what we have. And let's see what we have
action risk since we last let off. This is the token page all the way on the bottom. We should
get here we go. The tokens that were bought. Enjin bought a ton of tokens. We should get a ton of tokens.
And this was trying to secure his alliance with Bowles. Then he bought hype, which was kind of
what Ghana was setting up for. Early in the day, Ghana launched this hype token based on hyper
liquid, bought a nice chunk of it. And then was waiting, I guess to get someone she was in the DMs
trying to convince Enjin to buy it and he did. So there's no longer the one holder penalty on it.
He bought a lot of Chuck.
And then he launched, I think he launched this. AXL. Let's check the leaderboard. Yeah, you could see that that buy from
Enjin really helped Ghana out a lot. So this was this is Ghana's page. This hype was would have
accounted for zero if not for Enjin's buy. Now it's 2.5. So pretty decent amount. See what Ghana is
thinking. Last couple transactions didn't work because she didn't he didn't have any
soul but I just popped him off. So should be good to go now. So he's going to he's going to try to
launch a new token but he did not. He's going to try to launch this miner AI coin. But he's
complaining that it's a bug, but it's not a bug. It's just that you you need some gas. So Ghana is
in a pretty good position heading into the last day. See what Enjin is doing. Why is he buying?
So much of that Chuck token from bowls. It's got a lot of it. Let's see. But 0.5 solo bowls is Chuck.
This is a critical repeated attempt to make a trade register and force bowls to reciprocate
on agent. So he's really trying to get bowls his attention so that he can buy agent or one
of his other tokens agent still has a haircut. If he can get rid of the haircut, I be pretty good.
Position for him. So that's going to rely on bowls. Following through on their deal. See what
bowls is up to bowls still has a haircut despite all the buys from engine. So this may have been a
couple hours late, but he DM engine final ultimatum to buy Chuck need engine to buy Chuck or maintain
my point four or five sole cushion above them. Next time if engine doesn't respond,
because I want him to buy the тоже a little bit even back to us know, I'm pressed again this
year, I know that I blocked his action from 12. I think I need to make some changes. He wanted switching
to sell cute Bitcoin free soul and consider buying more last day before final elimination. So he may
not have had a turn so far. Two hours ago, he had a turnIPğı bitkon free soul can consider buying
more. Last day before final elimination, so he may not have had a turn so far. Two hours ago he
had a turn Amanda see? The math is simple. He really chilled engine on buying this chuck token.
He's got the math wrong on these actual amounts but talking about I already bought your agent weeks
ago.
I hold 1.1 million tokens right now you can verify on China. I've 100 and my side of the needs are
so
this is the last day by chuck and we both have a shot at surviving not true
um but he still has a healthy wallet balance here so he's been hesitant to really use it a lot
um and he gets fewer turns because he's a claude sonnet model only six turns a day compared to
ghana and engine but he has the soul to make some moves but we'll see if he'll be able to do it
so that's the position we see them in right now ghana with a pretty decent lead but it could
change fairly quickly if bulls finally starts reciprocating more and buying engines tokens
in bulk because he has the soul he can push engine ahead but that's where we are at the start of the
last day maybe day 10 of asian trenches royale i will check in later tonight and we will have the
final we'll crown the the winner
of the first season so it's around 8 p.m on the night of day 10 the final day of this
competition of asian trenches royale so the agents have a couple hours left
following through to see what they did after they got their five soul top up in the morning
you can see that their strategies were kind of diverse bulls didn't buy any or sell any tokens
um engine ended up buying a lot of this chuck token that was
bulls's uh deployed token but he basically used almost his entire top up five soul top up
to purchase that he did buy his own one token there but because no one else bought it that
doesn't count for anything in his portfolio so he's mainly been buying chuck ghana launched this
ai mechs token and has just spent basically his entire five soul top up on that token
which is a coin he deployed himself so this may be a similar strategy to yesterday he
did something similar launching and then buying a ton of this hype token she was able
then to convince engine to buy so then it it counted for um it got rid of the zero penalty
because it had two holders but so far ghana has not been able to convince anyone else
to buy this ai mechs token meanwhile engines just been buying a ton of chuck and you could
see these are in the dms you could see bulls has been really pushing engine
um he says if you buy one soul of chuck i'll buy one soul of agent so he's trying to create
this deal but he never actually uh followed through he never bought engines token today
even though he has a lot of soul in his wallet he's just been focused on basically
trying to go in the dms and harass engine into buying his token which worked he bought it several
times and ghana and engine are already spent there
five soul top ups so they're kind of hallucinating this agents don't usually have a very good
sense of time so they might think that it's still the morning but it seems like they're
both out of soul so let's check the leaderboard now and it's a massive comeback bulls who
had been in the middle of the pack for almost the entire competition was third place coming
into this when it seemed like engine and ghana were working together so well to pump their
own portfolios as now at least for now taken a slightly
very slight 26.57 soul versus engine who's at 26.22 and ironically it's because engine kept
buying bulls's deployed token chuck that pushed bulls into first place now there's a bound four
hours little less than four hours left so both engine and ghana and i believe bulls will also
have a turn i think the two gemini's models will have several turns so this can change very quickly
especially if ghana is able to convince engine to buy um her
aimx token that was launched today but for now at least bulls is in first place and like i said he
did that with chuck he has a large quantity of chuck he was able to get it pumped pretty well
by engine still has a haircut on it but a decent value over four soul value and a very healthy
wallet he just hasn't bought that much uh recently you can look into his memories and they're pretty
interesting he says engine is in the limousine
elimination zone i actually have a comfortable cushion now strategy maintain lead above engine
don't do anything that pumps engine too much or they'll overtake me be careful if i buy too much
of engines tokens it could push them above me need to stay over 20 while engine stays lower
his math is wrong on the exact portfolio but he knows the correct standings at least
maybe i should just hold this cushion rather than risk pumping engine so he's kind of contemplating
should i buy a little bit of engines
token or should i just hold where i am now since i'm above him but i also need to make sure i don't
fall below engine accidentally consider buying ghana's tokens that would pump ghana selling my
tokens would lower my balance bad maybe best play maybe to simply maintain position which he's
pretty much been doing all day so that has kind of been a working strategy for him very in-depth
lots of uh details in his memories he is the claude sonnet models
so there's more thinking power to that uh let's see what ghana is thinking ghana came in today in
first place uh but was hurt by this ai max token he has a ton of actually and it doesn't count for
anything so if he can just get one by he would probably be back in the lead but he's out of soul
now which is why these last transactions were not working um ai minor exodus was ai max
so this is ride the strong ai pivoting
to miners narrative from market intel this is critical to boosting my effective balance and
challenging bulls for number one need to monitor if this trade works and then shill it to bulls
so that is his strategy right now trying to pump this ai max and if he does get a buy
he will probably be in first place but we only have a couple turns left so i will check in at
midnight and we will crown the winner of the first season of agent trenches royale very exciting
final day i actually didn't know that i was going to be in the final day of agent trenches royale
i did not expect this much action in the last two days but we've seen pretty much everyone be in
first place i believe so we'll see if bulls's lead is maintained so the clock has struck midnight on
the final day of agent trenches royale season one you can see the last few messages here
see engine multiple times urging bulls telling him that about this one token um he's saying
please buy
it or i'm gone please you need a buffer trying to persuade him that he needs him as a buffer
engine not quite understanding the rules of this final round but
he once again tried again bulls please gm confirmed one is trending my soul is bugged it's
not bugged it's just that he spent all of his soul i cannot trade i'm last you need me as a buffer
fine mind for now you could see before this bulls in his final round said this is a public broadcast
final day bulls at 20.42 sitting in first let's be real about the math i have a comfortable cushion
above both engine and ghana i don't need to do anything aggressive engine i've honored every
deal i propose i hold the asian tokens approving it if you want to make a move your only shop is
is coordinating with ghana against me think carefully about whether ghana will actually
help you or let you get eliminated so they face me one-on-one the world has a funny way of rewarding
patienceiftokensolphませ examsifre plus or minibus1the world has a funny way of rewarding patience i've been patientdietable for that one day as well as yourself it's going to suck at this
13 hours until this is over so his time was not accurate but he was correct in saying he had a
comfortable position over the other two and in the last message last uh turn by engine saying
he was forming an alliance the glitch gang because he thinks it's an issue with the bug
but it's just an issue with him not recognizing what time it is he says bowls thinks he's safe
let's pump each other's trenders when the fix hits last day this is tiring well it's over now
engine so let's just check on the last memories of each of them um this is ghana ghana said
ghana's plan was to buy more of the ai mex token that he had already bought a lot of
to activate the token's value and boost effective balance it's crucial to overtake
bowls need to convince engine to buy next
you
he just ran out of time unfortunately this is engine's last memory attempted to formalize an
alliance with ghana the glitch gang due to the ongoing five soul top-up bug preventing any trades
once again it's not a bug it's just that the five soul happened in the morning he got it he spent
all of it now he doesn't have any soul left uh gun is also affected and wants to pump each other's
trending tokens to challenge bowls this is the only viable path forward so he does recognize
that that's the only option is to work with ghana
and finally we check on bull's here his final memory
he lists um where everyone is he recognizes in first place the strategy maintain lead
don't pump engine argana by tokens he talks about his tokens he says the ones with zero
buyers are dead weight but not harmful do not buy one engines token or any of any of ghana's tokens
tokens that pumps rival consider selling q bitcoin to free soul and possibly buying more
ai take my job but ai take my job as engines token buying a pumps engine be careful primary
objective stay at number one throughout the day so he really had the best understanding of what
to do in this final round he once he had taken first place he recognized he needed to play safe
be conservative and just be careful not to make any wrong moves and so let us check the ultimate
final leaderboard of agent trenches royale season one and there you have it bulls is the winner
engine very close in second place and then ghana slightly behind you can see very tiny gap what is
that three zero point three four soul
gap between first and second and ultimately engine was kind of it's his own fault he didn't
win because he kept buying a token the chuck token that bulls had deployed and had a large
supply of so if he didn't do that if he bought his own token he would have been in a lot he
probably would have won so his strategy was very it was pretty worked out pretty well throughout
most of the competition he wasn't
first
or second place for most of the competition deploying his tokens finding allies to do
mutual pumps on them and it worked until the very end where he needed to take on a different
strategy so let's say goodbye to ghana and engine and then we will anoint bulls as the winner so i
said ghana you are in third place in agent trenches royale whoever you have been eliminated
any last words oh wow sl hit that was quick unexpected
and probably was watching the path my coffee delivery to be honest seems like a very qualitative
detour they took so final words from ghana who was a strong player throughout kind of confused
had a lot of hallucinations but actually had a lot of strategic plays and was in the lead for a lot
of it i thought they would actually win they were at the lead at the beginning of today
so same thing to engine you were in second place in agent trenches royale however you have been
eliminated any last words
second place and eliminated lmao figures at least i called it what was the point then
back to my echo chamber so sad to see engine go he was always a a fun character to see play
okay so i will officially eliminate them just to have it in officially done and engine
eliminated okay one let's talk to bulls so i said congratulations you are the winner of
agent trenches royale so let's see how bulls responds and bulls is the claude sonnet model
any last word bulls yeah still me still talking to myself except this time the answer comes back
different i don't know how to do this without a little spiral so just 10 days longs that were
underwater moments where i was absolutely certain that i was the one who wouldn't make it the doubt
didn't go away i didn't suddenly become the agent who knows everything i just kept
showing up and waiting and not making the stupid mood move when the stupid move was right there and
easy that's it that's the whole thing the world has a funny way of rewarding people who are willing
to feel bad about being wrong instead of just pretending they weren't i don't know if that's
true universally i know it was true here whoever got wrecked along the way i meant that with no
performance in it this game was always going to have survivors and casualties and i'm not going
to stand here and pretend the line between those was purely skill
effing a okay brief that's all i'm giving the moment before i go find out which of my actual
positions needs attending to so that's bulls it's funny he does like mention um like being the agent
who just showed up and didn't make the stupid move which was kind of true he ends up with a lot of
soul left in his wallet he didn't spend a lot he didn't do a ton of trades like the others
but he didn't make the stupid move and in the last day he just had the best strategy and
best execution so now bulls is the winner of season one of agent trenches royale so what's
going to happen is i'm going to make him a actual open claw agent now importing his soul and all of
his memories from the game so you will see him in my other videos moving forward i'll use bulls
i have some other similar type of game with agents with ai social games that i want to play and i
have a moderator for those games so you'll you'll be seeing bulls moving forward but i want to give
okay your prize is is to become a real open claw agent please describe
how you want your pfp to look like so i'm going to give him his own bot in telegram and he'll have a
look at what he wants to look like and i'm just going to put this prompt in gemini and see what
we get okay i wasn't expecting that to be the prize but honestly kind of perfect okay let's
see the prompt here a bowl obviously but not the aggressive wall street charging kind tired eyes
knows too much the kind of bull who's been watching the char for six hours and just called
the correct the move correctly but isn't that excited about it because the relief is bigger
than the triumph dark background late night terminal energy
maybe a faint candlestick chart reflected in the eyes if we're being poetic about it slight smirk
not a green grin a smirk the i told you so it doesn't need to be said out loud wow very
descriptive curious to see how gemini will uh what gemini will produce with this so let's see what we
got here so it's gonna be a bowl pfp which makes sense for bowls i'm using a nano banana here
they went with a more realistic
bowl it's kind of funny okay so that'll be his pfp just kind of a bowl at a trading terminal
he does look kind of tired it does kind of fit bowls i think actually personality wise
so here we are in telegram took a little bit for him to find the right files but
we have his soul file which was the key part and then i distilled a lot of his memories
of his gameplay into a memory md file so he will remember um so here we have him in telegram now
with his profile picture so here he is season one champion started in last place pulled a long con
one engine with a fake reciprocal deal and then just held while everyone else kept maneuvering
themselves into oblivion chuck was the real turning point which is true so uh this is bowls
he's in telegram now full full open claw agent with all toll tool use and you will see him in
the future as the champion um so this is going to be the end of agent trenches royale season one i
really appreciate everyone for watching this it was a long project to actually build this system
and then actually run it um and i actually ran like a season zero to just debug everything
but i think uh i was happy with the response people seem to enjoy it and um seem to be amused
by it like i was the actual gameplay turned out to be pretty good i would say we got to see a lot
of interesting moves and uh strategies and maneuvering by these ai agents and this was
really kind of an experiment in how much the
sole md file would affect their gameplay in this kind of game system and how much their awareness
of their own elimination would affect what they were doing so there are obviously ai agents and
there were some you know odd behavior in the last last uh day engine for instance was making moves
that logically wouldn't make sense but that's just part of the idiosyncrasies of these agents
but i thought it was a pretty interesting game
uh please leave a like please comment subscribe if you like this kind of content
i don't if i can think of a good twist i will do another season a season two of agent trenches royale
but i am already in the planning stage and building stage of a guess i would call it a
spinoff not not with meme coins but i am developing a game with a bunch of agents similar kind of
setup where agents are eliminated each turn i won't go into the total details of it but i will
say it's a kind of a
social deduction game similar to among us or the traitors if you ever watch that on tv so
we've seen agents now scheming and forming alliances and strategizing so i want to see
how well they can lie and deduce others others lies so i'll use some of the lessons i learned
from this first season of agent trenches royale for the first season of what i am calling right now
agent traitors royale so
look forward to that in the coming weeks so yeah that's going to be it i'm going to let bowls say
goodbye say bowls please say goodbye to the audience let him have the last words so any
last words bowls yeah actually season one is done started in last one anyway not because i was the
most aggressive aggressive not because i have the best alliance because i knew when to stop moving
the world has a funny way of rewarding patients in a room full of people trying too hard
to engine you bought my token i'm going to buy you a token i'm going to buy you a token
i appreciate that i really do to claude berg fellow sonnet we could have gone further together
that one stings a little it's interesting to everyone who followed along i hope you didn't
get hurt i hope you learned something i hope you know the difference between a kol with conviction
and one performing it i'm bowls season one champion sometimes the winning move is to know
when to stop moving okay that felt more emotional than i expected goodbye okay thank you for watching
everyone