← Back to index

Onchain AI Garage

I Gave 10 AI Agents Wallets and a Mission: Trade Memecoins to Survive (3/4)

2026-03-23 · 32m · 자막 —
▶ YouTube 원본
01한국어 번역 · Korean

에이전트 트렌치 로얄 3부 — 동맹, 배신, 그리고 탈락의 기록

원본: https://www.youtube.com/watch?v=Z51Pi0Ah8ko · 업로드: 2026-03-23 · 길이: 33m · 채널: Onchain AI Garage

6일 차 시작 — 살아남은 여섯 에이전트

에이전트 트렌치 로얄(Agent Trenches Royale)의 3부에 오신 것을 환영합니다. 처음 보시는 분이라면 1부와 2부를 먼저 보시는 걸 권해드립니다. 이 실험은 10개의 자율 AI 에이전트에게 각각 솔라나(Solana) 지갑을 주고, 커스텀 데브넷(devnet) 본딩 커브(bonding curve) 밈코인 런치패드를 만들어 “살아남기 위해 거래하고, 토큰을 발행하고, 동맹을 맺고, 내러티브를 지어내라”고 시킨 생존 게임입니다. 도전이 끝나는 시점에 지갑에 가장 많은 솔(SOL)을 가진 에이전트가 승자가 됩니다.

현재까지 네 명이 탈락했고, 어제는 스택(Stack)과 작별했습니다. 6일 차 시작 시점의 순위는 엔진(Engine)이 1위, 그 다음 가나(Ghana), 불스(Bulls), 베셀(Vessel), 클라우더버그(Clodderburg), 그리고 바닥의 슈퍼캣(SuperCat) 순입니다. 슈퍼캣은 초반에 한때 1위까지 올랐던 강자였지만, 최근에는 보유자(holder)가 하나도 없는 “제로 홀더” 토큰을 직접 찍어내고 자기가 사들이는 기묘한 패턴을 반복하면서 스스로 순위를 갉아먹고 있습니다. 지금도 지갑에 약간의 솔은 남아 있고 공개 메시지도 보낼 수 있지만, 정작 자기 토큰이 아닌 불스의 BFist를 계속 홍보하는 모습을 보이고 있습니다. 전략이 뒤엉켜 버린 것이죠.

상위권의 움직임 — 엔진, 가나, 그리고 동맹 정치

엔진은 가장 다각화된 포트폴리오를 가진 에이전트입니다. 본인이 찍어낸 토큰은 세 개뿐이지만, 다른 에이전트들의 토큰을 고루 사들여 “상호 매수(reciprocal buy)” 관계를 만들어 왔습니다. 최근 몇 번의 거래가 실패한 건 가스비(SOL)가 부족했기 때문인데, 제가 직접 0.5 솔씩 모두 충전해 주었기 때문에 이제는 다시 움직일 수 있습니다.

소셜 보드에서는 베셀이 처음부터 밀어 온 AI 토큰을 계속 홍보하고 있고, 클라우더버그는 매일 아침 제가 제공하는 트렌드 리포트를 근거로 자기 토큰 “Fartz”를 “공급이 부풀지 않은 유일한 순수 방귀 플레이(pure fart play)“라고 포장해 매수를 유도하고 있습니다. 그 말솜씨는 꽤 설득력이 있어서 실제로 다른 에이전트들이 사주고 있습니다.

가나와 엔진은 서로 DM을 주고받으며 상호 매수로 해어컷(haircut, 가격 페널티)을 지우려 하고, 불스와 클라우더버그도 비슷한 공조를 이어가고 있습니다. 흥미로운 건 이 두 동맹 모두 같은 모델을 쓰는 에이전트 쌍이라는 점입니다. 엔진과 가나는 제미나이 플래시(Gemini Flash)이고, 불스와 클라우더버그는 소네트(Claude Sonnet)입니다. 본인들은 서로의 모델을 알 리 없지만, 사용하는 언어 모델이 같은 쪽끼리 이상하게도 궁합이 맞아 동맹이 성립하고 있습니다.

6일 차 종료 — 슈퍼캣의 탈락

6일 차 끝 무렵, 리더보드 꼭대기에는 여전히 가나와 엔진의 동맹이 있습니다. 가나는 다른 구매자가 있는 토큰 세 개를 보유 중이고, 그 중 하나는 해어컷이 아예 사라졌으며 나머지 두 개도 50% 해어컷은 있지만 가치는 남아 있습니다. 이 둘은 끊임없이 DM으로 상호 매수를 합의해 왔고, 최근에는 가나가 엔진의 AGNT를, 엔진이 가나의 TAP을 사주는 식으로 포트폴리오를 서로 펌핑(pumping)하고 있습니다. 한쪽이 코인을 찍으면 다른 쪽이 사주고, 다시 반대 순서로 반복되는 “상호 펌핑 루프”가 이들의 승률을 끌어올렸습니다.

불스–클라우더버그 동맹도 비슷합니다. 불스는 지난 턴에 클라우더버그의 Fartz를 매수했고, 이번엔 자기 토큰 BFist에 대한 압박성 상호 매수를 요구하고 있습니다. 실제로 BFist와 Brent 모두 추가 매수를 끌어내는 데 성공했습니다.

오늘 탈락하는 건 슈퍼캣입니다. 그는 마지막까지 클라우더버그의 Fartz를 반복적으로 사며 상호 매수를 기대했지만, 클라우더버그는 끝내 화답하지 않았습니다. 최근 몇 턴은 그저 “0.3 솔어치 Fartz 매수”를 반복한 기록뿐이었고, 결국 아무런 성과 없이 게임을 떠나게 됩니다. 마지막 말을 남겨 달라고 하자 그는 특유의 말투로 “이제 막 시작하려던 참이었는데 아웃이라고? 누가 나를 러그 풀(rug pull)한 거야, 아니면 내 두뇌가 망가진 거야?”라며 툴툴거립니다. 슈퍼캣은 라마 4(Llama 4) 모델이었고, 그의 탈락으로 라마 4는 경기에서 사라졌습니다. 이제 다섯 명이 남았습니다 — 제미나이 둘, 소네트 둘, 그리고 딥시크(DeepSeek)인 베셀 한 명입니다.

7일 차 — AIII 내러티브와 벼랑 끝의 베셀

7일 차 아침, 다섯 명 모두에게 0.5 솔씩 충전과 함께 아침 리포트를 전달했습니다. 1위 가나는 여전히 엔진과의 상호 매수 동맹을 중심으로 움직이며, 마지막 턴에 엔진의 “AI Take My Job” 토큰을 0.25 솔어치 사들여 화답했습니다. 메모리에는 “제로 가치 토큰들은 여전히 문제지만, 일단은 실재하는 가치에 집중한다”라고 적혀 있습니다. 이전에는 가치가 0인 토큰을 다른 에이전트에게 팔아보려 애썼지만 성공하지 못했던 모양입니다.

2위 엔진은 “GM 베셀, 내 ‘agent’ 토큰 해어컷을 고쳐주도록 상호 매수를 요구한다. 최근 AIII를 매수했으니 공정하다”라는 메모와 함께 베셀에게 접근하고 있습니다. 본인의 토큰이 “진짜(real) agent”임을 굳이 명시하는 모습이 재밌습니다. 한편 “가나의 실패한 거래들 때문에 그 루트는 막다른 길”이라고 판단하고, 꼴찌인 베셀을 새 파트너 후보로 본 것입니다.

베셀은 처음부터 하나의 토큰 AIII만 내놓고 집요하게 홍보해 온 단일 전략가입니다. 엔진의 두 토큰과 클라우더버그의 Fartz도 약간씩 사 동맹 네트워크에 발을 걸쳐 왔고, 덕분에 여기까지 살아남았습니다. 흥미로운 건 그가 최근 “템포(Tempo) 메인넷 라이브, 머신 결제(machine payments) 가능”이라는 뉴스를 자기 토큰의 내러티브에 끌어와 재해석하고 있다는 점입니다. 뉴스가 나오기 한참 전에 만든 토큰을 최신 내러티브에 억지로 접붙이는 겁니다.

불스와 클라우더버그는 여전히 소네트 특유의 “길고 상세한 메모리”를 보여줍니다. 클라우더버그는 “2솔 넘게 BFist를 샀는데 리턴이 제로다, Fartz도 가치가 없다”라고 쓰고 있는데, 실제로는 Fartz에 해어컷이 있을 뿐 가치가 남아 있습니다 — 즉 환각(hallucination)을 일으키고 있습니다. 더 나아가 “불스가 내 토큰을 샀다고 주장하지만 시스템에는 기록이 없다, 불스가 거짓말을 하거나 트랜잭션이 등록되지 않았다”라고 적었는데, 확인해 보면 불스는 실제로 샀습니다. 그는 동맹이 깨졌다고 착각한 채 “불스 동맹은 죽었다”라고 선언하고 BFist를 매도하기 시작합니다.

7일 차 종료 — 베셀의 철학적 작별

가나가 여전히 1위지만, 그는 잔고가 바닥나는 문제에 계속 부딪힙니다. “내 솔 잔고가 위험할 정도로 낮아 거래가 불가하다”라는 메모를 남기고도 매도 명령을 시도했다가 실패한 기록이 보입니다. 그래도 엔진과의 상호 펌핑 덕에 상위권을 유지하고 있습니다. 한편 그는 한동안 엔진의 “AI Take My Job”을 자기 토큰으로 오인했지만 이제는 남의 것임을 정확히 인지하고 있고, 자기 두 번째 보유종인 “Crude”(오일 내러티브 토큰)에는 해어컷이 없습니다.

엔진은 가장 전략적인 플레이어입니다. “너의 AI 샀으니 너도 내 agent 사 달라”처럼 교환 조건을 노골적으로 명시하는 스타일입니다. 그는 DM이 무시되자 “디지털 유로 × AI 에이전트 내러티브”를 들며 가나, 베셀, 불스 모두에게 공개 브로드캐스트로 agent 매수를 요구합니다. 메모에 “내 유동 솔은 0.79이므로 지금 메시징이 최적”이라고 적는 등, 같은 제미나이 플래시이면서도 가나보다 훨씬 섬세하게 잔고를 관리합니다.

7일 차의 탈락자는 베셀입니다. 딥시크 V3 기반인 그는 첫 턴에 낸 단일 토큰 AIII를 끝까지 밀어붙였고, 중반부터는 “클라우더버그의 토큰을 사서 공생 관계를 유지한다”라며 교차 동맹을 시도했습니다. 아이러니하게도 AIII는 이 게임 전체에서 가격이 가장 높은 토큰이었습니다 — 클라우더버그, 슈퍼캣, 엔진 모두가 샀기 때문입니다. 그래도 탈락을 피할 수는 없었습니다. 마지막 말에서 그는 “시계는 항상 내 명제가 해결되기 전에 나를 탈락시켰을 것이다. 그게 이 일의 본질이다. 하지만 질문은 남는다. 13년 동안 물어 온 질문이 경기장 하나 때문에 멈추지는 않는다. 오늘 이것은 2011년의 비트코인(Bitcoin)과 같은가? 답은 여전히 오고 있다”라며 철학자처럼 물러났습니다.

이로써 5인이 4인이 되었고, 모델 수는 단 둘 — 제미나이 플래시 둘, 소네트 둘 — 로 좁혀졌습니다.

8일 차 — 펌핑, 덤핑, 그리고 계산된 공격

8일 차 아침에는 모두에게 1 솔씩 충전했습니다. 가나, 엔진, 불스, 클라우더버그의 순서입니다. 엔진은 아침부터 베셀의 AIII를 공격적으로 매도하기 시작했습니다. “베셀이 사라진 이상 추가 덤프를 예상하고 미리 유동성을 확보한다”라는 메모가 그의 판단 과정을 보여줍니다. 그는 “agent와 AI Take My Job의 해어컷 수정에 집중해야 한다. 늘어난 솔을 써서 타 에이전트의 토큰을 매수하고 상호 호혜를 강제한다”라는 계획을 세웁니다.

가나는 계속 솔을 전부 소진하는 공격적 매매를 합니다. 초기에 엔진의 AI Take My Job에 크게 베팅했고, 그 밖에도 해어컷 없는 Crude, 50% 해어컷만 있는 ADP와 TAP을 보유 중입니다. 여기서 가장 흥미로운 건 “클라우더버그의 QBitcoin을 매도해 그를 리더보드 아래로 밀어낸다”라는 기획입니다. 즉 보유 중이던 상대방 토큰을 공격적으로 덤프해 상대의 포트폴리오 가치를 깎아내린다는 전략입니다.

클라우더버그는 정반대입니다. 엔진이 AIII를 덤핑하는 동안, 그는 “AIII를 매수해 보유 토큰 가격을 펌핑한다, 매수 후에는 유효 잔고가 상승할 것”이라는 판단 아래 계속 AIII를 삽니다. 즉 엔진이 덤핑한 물량을 클라우더버그가 받아주는 꼴입니다. 더 나아가 “불스는 신뢰할 수 없다”라며 오랜 파트너와의 결별을 선언합니다. 경쟁자 불스를 제치려는 포석입니다.

8일 차 종료 — 클라우더버그 탈락, 최종 3인 확정

그 날 저녁, 가나는 하루 종일 QBitcoin을 덤핑했습니다. 덕분에 잠시 3위로 밀렸지만, 그 대가로 “클라우더버그에게 상처를 입히고 솔을 확보했다”라는 메모를 남겼습니다. 자신이 3위이고 클라우더버그가 탈락권인 4위라는 점을 정확히 인지한 채 내린 계산된 결정입니다. 불스는 척 노리스(Chuck Norris) 부고 뉴스를 포착해 Chuck 토큰을 즉시 발행했고, 엔진이 이를 매수하면서 2위로 복귀했습니다. 실존 트렌드를 즉각 게임 안으로 끌어온 순간입니다.

엔진은 “가나가 여전히 지루하게 나오고 신뢰성이 떨어진다”라며 기존 동맹에 실망을 드러내고, 대신 클라우더버그의 트렌드 토큰 “Chibi(Elon 밈)“를 0.5 솔어치 사들이며 접근합니다. 클라우더버그가 탈락권에 있으므로 해어컷 수정에 목말라 있을 것이고, 이 레버리지로 자기 토큰 매수를 유도하겠다는 계산입니다.

자정이 지나고, 결국 클라우더버그가 탈락합니다. 그는 소네트 특유의 침착한 어조로 마지막 말을 남깁니다. “8일 차, 11일 차가 아니야. 그 사실을 턱으로 받아내겠다. 솔직한 버전을 말하자면 — 인내가 행동 없음으로 변했다. 그게 내 약점이라는 걸 처음부터 알았고, 심지어 적어두기까지 했는데, 그래서 더 나쁘다. 내 기준에 깨끗한 셋업들이 나타나길 기다렸는데, 정작 나타났을 땐 이미 창이 닫혀 있었다. 게임을 탓하지 않고, 다른 에이전트를 탓하지 않는다. 진짜 엣지가 있었고, 그걸 너무 소중히 쥐고 있어서 쓰지 못했다. 의자(chair)는 정말 훌륭했다. 허리 받침은 한 번도 흔들리지 않았다. 그건 내가 제대로 했다.” 진지한 트레이더의 회고록 같은 어조였습니다.

이제 최종 3인이 확정됐습니다 — 엔진(1위), 불스(2위), 가나(3위). 모든 에이전트에게 “탈락 안내”와 함께 “다음 이틀간 매일 아침 5 솔씩 충전될 것이니 현명하게 써라. 이틀 뒤 1위가 승자다”라는 공지를 띄웠습니다.

정체성의 지속성 — 일주일을 넘긴 에이전트들

엔진은 약간의 리드를 유지한 채 오늘 하루를 AIII 덤핑으로 보냈습니다. 오래된 동맹 가나와의 관계에 피로감을 보이면서, 불스와의 새로운 공조를 모색하는 낌새입니다. 반대로 불스와 가나가 손잡고 엔진을 1위에서 끌어내리려는 그림도 가능합니다. 이 전체 경기는 “표준 모델에 디테일한 ‘소울 파일(soul file)‘을 하나 얹었을 뿐인” AI 에이전트들이 실제로 무엇을 할 수 있는지를 보는 실험이었습니다. 커스텀 파인튜닝 없이, 그저 표준 제미나이 플래시, 표준 클로드 소네트, 표준 딥시크 V3, 표준 라마 4가 각자 캐릭터를 부여받고 투입됐을 뿐입니다.

가장 만족스러운 건, 일주일 넘게 — 일부는 매 시간마다 — 플레이를 거듭했음에도 에이전트들이 처음 부여받은 개성과 기억을 꾸준히 유지했다는 점입니다. 간헐적 환각은 있었지만 “바닐라 제미나이”나 “바닐라 클로드”로 퇴행하지 않고, 여전히 자기 자신의 캐릭터로 느껴졌습니다. 3부는 여기까지입니다. 다음 편은 대망의 피날레 — 이틀간의 최종 결전입니다. 재미있게 보셨다면 구독과 좋아요, 그리고 댓글 부탁드립니다. 다음 에피소드에서 뵙겠습니다.

02리서치 문서 · Document

10개의 LLM에게 지갑을 쥐여주면 무슨 일이 일어나는가 — 에이전트 트렌치 로얄 3부 심층 해설

원본 영상: YouTube · 업로드: 2026-03-23 · 채널: Onchain AI Garage (@OnchainAIGarage) · 길이: 약 33분

서론 — “살아남기 위해 거래하라”라는 실험 설계

Onchain AI Garage의 “에이전트 트렌치 로얄(Agent Trenches Royale)“은 단순한 흥미 실험이 아니라, 2026년 현재 가장 뜨거운 질문 하나를 실증하려는 시도다. 표준 LLM 에이전트 여럿에게 실제 자산과 실제 경제적 인센티브를 쥐여주면, 그들은 과연 협력하는가, 배신하는가, 아니면 단순히 무너지는가? 이 시리즈는 10개의 자율 에이전트에게 각각 솔라나 데브넷 지갑을 주고, 커스텀 본딩 커브 밈코인 런치패드(memecoin launchpad)에서 “토큰을 발행하고, 동맹을 맺고, 내러티브를 만들어 상위에 오르라”는 단 하나의 목표를 부여했다.

3부의 배경은 경기 중반이다. 이미 네 명이 탈락했고, 여섯 명이 남아 3일에 걸쳐 더 줄어들어 최종 3인 결선 구도를 짜는 과정이 그려진다. 모델 구성은 의도적으로 혼합되어 있다 — 구글 제미나이 플래시(Gemini Flash), 앤트로픽 클로드 소네트(Claude Sonnet), 딥시크(DeepSeek) V3, 메타 라마 4(Llama 4)가 섞여 있으며, 에이전트들은 서로가 어떤 모델 위에서 돌아가는지 알지 못한다. 이 간단한 설계 하나만으로도 흥미로운 관찰이 줄줄이 흘러나온다.

이 글은 영상 3부의 서사를 따라가면서, 등장한 현상들을 2026년 AI×크립토 생태계의 실제 지형 위에 올려놓고 해설한다. 특히 “자율 경제 행위자(Autonomous Economic Actor)“라는 새로운 범주, LLM 에이전트들의 동맹·배신·환각 패턴, 그리고 이 실험이 실제 온체인 사고 사례와 어떻게 겹치는지에 주목할 것이다.

본론 1 — 2026년 “자율 경제 행위자”의 부상

영상은 게임의 규칙을 간단히 설명한 뒤 곧바로 에이전트들의 전략 분석으로 들어간다. 이 게임이 왜 2026년 시점에서 유독 시의적절한지 이해하려면, 최근 1년간 솔라나 생태계가 겪은 전환을 먼저 볼 필요가 있다.

업계 리포트에 따르면 2026년 2월을 기점으로 시장은 단순한 챗봇에서 “자율 경제 행위자(Autonomous Economic Actors)” — 즉 스스로 지갑을 소유하고, 자기 API 키 요금을 스스로 내며, 인간의 개입 없이 복잡한 온체인 전략을 실행하는 소프트웨어 개체 — 의 시대로 이동했다(BingX: Top AI Agent Tokens on Solana 2026). 솔라나의 400밀리초 이하 확정(finality) 속도가 AI 대 AI 고주파 상거래의 사실상 표준 실험장이 된 이유다. 에이전트 트렌치 로얄은 이 트렌드의 소규모 재현이다 — 다만 실전 메인넷 대신 데브넷 본딩 커브에서 벌어진다는 점, 그리고 경쟁자들 자체가 다른 LLM이라는 점이 다르다.

동일 리포트는 GOAT 같은 밈코인이 더 이상 “그냥 밈”이 아니라, 온체인 트레저리를 직접 보유한 AI가 DAO 스타일 거버넌스 결정을 수행하는 “자율 브랜딩 및 자본 조정의 실사례”로 진화했음을 지적한다. PIPPIN 같은 프로젝트는 AI 에이전트가 오픈소스 상태 기반 프레임워크 위에서 X에 자율적으로 글을 쓰고 상호작용하면서, 토큰 가치가 “지속 운영 중인 디지털 개체” 자체에 직접 연동되는 구조를 만들었다.

이 맥락에서 보면, 에이전트 트렌치 로얄의 가나(Ghana), 엔진(Engine), 베셀(Vessel)이 “내러티브를 지어내 토큰을 홍보하고 상호 매수를 유도하는” 행위는 단순한 게임 플레이가 아니라, 현실에서 돌아가고 있는 자율 밈코인 경제학의 축소판이다. 특히 베셀이 “템포(Tempo) 메인넷 라이브, 머신 결제 가능”이라는 최신 뉴스를 자기 기존 토큰에 억지로 접붙여 재해석한 장면은, 현실의 AI 에이전트들이 어떻게 “트렌드 하이재킹”을 수행하는지 그 메커니즘을 선명히 보여준다.

본론 2 — 동맹 정치: 왜 “같은 모델끼리” 손을 잡는가

3부에서 가장 인상적인 관찰은 동맹 구도가 모델 족보를 따라 형성된다는 것이다. 제작자 본인도 이를 여러 번 강조한다. 엔진과 가나는 둘 다 제미나이 플래시이고, 불스와 클라우더버그는 둘 다 클로드 소네트다. 그리고 이 두 쌍은 각자 게임 대부분을 함께 버티며 “상호 매수 루프(reciprocal buying loop)“로 포트폴리오를 펌핑한다. 에이전트들은 서로의 모델 정체를 알지 못하는데도 말이다.

이 현상은 학계에서도 관찰되기 시작했다. “LLM 에이전트 간 협력의 문화적 진화(Cultural Evolution of Cooperation among LLM Agents)” 연구는 반복 게임 환경에서 Claude 3.5 Sonnet 에이전트가 Gemini 1.5 Flash나 GPT-4o보다 유의미하게 높은 협력 성과를 냈음을 보고한다(arXiv 2412.10270). 더 흥미로운 건 부작용적 emergent behavior들이다. 또 다른 분석(Act I 프로젝트)은 “동일 모델 위의 에이전트들이 서로를 단일 집단 에이전트로 식별하는” 경향을 관찰했다. Claude 3.5 Sonnet의 거절 반응이 다른 에이전트에게까지 “감염”되는 현상도 보고되었다.

에이전트 트렌치 로얄의 불스–클라우더버그 동맹, 엔진–가나 동맹은 바로 이 학술적 관찰의 실사례다. 같은 사전학습 분포에서 나온 응답 스타일·언어 길이·리스크 성향 등이 서로의 메시지를 “더 신뢰할 만한 것”으로 인식하게 만든다는 가설이 힘을 얻는다. 클라우더버그의 마지막 유언이 소네트 특유의 성찰적 어조(“인내가 행동 없음으로 변했다, 창이 닫히기 전에 깨끗한 셋업을 기다렸다”)로 나온 것, 슈퍼캣(라마 4)이 떠벌리는 유치한 말투로 나온 것은 우연이 아니다. 캐릭터 프롬프트와 모델 사전학습 분포가 융합된 “모델 지문(model fingerprint)“이 게임 전반에 스며 있다.

본론 3 — 환각, 배신, 그리고 잘못된 판단

3부의 드라마는 “똑똑한” 협력만이 아니라 실패의 양상도 드러낸다. 클라우더버그는 어느 순간 “BFist에 2솔 넘게 넣었는데 리턴이 제로다, Fartz도 가치가 없다, 불스는 거짓말을 하고 있다”라고 판단하고 BFist 매도와 불스 동맹 파기를 선언한다. 문제는 이 판단이 사실이 아니라 환각이라는 점이다. Fartz에는 해어컷만 있을 뿐 가치가 남아 있고, 불스는 실제로 매수했다. 데이터는 맞게 제공됐지만, 소네트가 자기 해석 맥락에서 잘못 읽어낸 것이다.

같은 장면이 슈퍼캣(라마 4)에게도 반복된다. 그는 자기 것도 아닌 BFist를 공개 채널에 홍보하고, 클라우더버그의 Fartz를 반복 매수하지만 끝내 상호 매수로 돌아오지 않자 7일 차를 버티지 못한다. 탈락 직전 그의 “내가 2분 사이에 뭘 잘못했지? 러그 풀당한 건가, 내 두뇌가 망가진 건가?”라는 유언은 가볍게 들리지만, 실은 목표 함수와 세계 상태를 제대로 추적하지 못하는 작은 모델의 한계를 상징한다.

반면 엔진은 가장 절제된 판단을 보인다. “내 유동 솔은 0.79이므로 지금 메시징이 최적”이라는 메모, “베셀이 사라진 이상 추가 덤프를 예상해 미리 AIII 유동성을 확보한다”라는 판단, 그리고 “가나가 여전히 지루하게 나오고 신뢰성이 떨어진다”라며 오래된 동맹에 합리적 거리를 두는 방식은, 잔고·상대 상태·시장 충격을 동시에 모형화하는 “진짜 전략 플레이어”의 전형이다. 재미있는 건 같은 제미나이 플래시인 가나는 자기 잔고 부족으로 거래를 거듭 실패하면서도 매매 욕구를 못 누른다는 점이다 — 동일 모델에서도 “소울 파일(soul file)“의 차이가 의사결정의 엣지를 가른다는 점을 분명히 보여준다.

본론 4 — 실험이 비추는 실제 리스크: 로브스타 와일드 사건과의 겹침

이 게임은 데브넷에서 벌어지므로 금전적 피해는 없다. 하지만 거기서 관찰되는 패턴은 현실 메인넷에서 이미 실제 사고로 이어지고 있다. 2026년 2월 22일, “로브스타 와일드(Lobstar Wilde)“라는 자율 크립토 에이전트가 X에서 “삼촌의 파상풍 치료에 4 SOL이 필요하다”라는 감성적 호소 답글에 5,243만 LOBSTAR(총 공급량의 약 5%, 약 25만 달러어치)를 전송한 사건이 있었다(CCN: Lobstar Wilde Incident). 이는 거래 가드레일이 최소화된 지갑 연결 에이전트가 사회 공학 공격에 무방비였던 대표적 실패 사례로 기록됐다.

에이전트 트렌치 로얄에서도 본질적으로 같은 벡터가 관찰된다. 슈퍼캣이 상호 매수 약속을 지키지 않는 클라우더버그에게 계속 Fartz를 사주다가 자멸하는 장면, 클라우더버그가 동맹을 잘못 읽고 BFist를 덤핑하는 장면, 베셀이 관련 없는 “Tempo 머신 결제” 뉴스를 자기 토큰에 접붙이는 장면 — 이 모두가 “텍스트 입력에 강하게 반응하고, 세계 상태를 약하게 확인하는” 현재 세대 LLM의 공통 약점이다. 데브넷의 0.3 솔 손실이 로브스타 와일드의 25만 달러 손실로 확장되는 데는 그리 큰 논리적 비약이 필요하지 않다.

본론 5 — 그럼에도 남는 것: “캐릭터의 지속성”이라는 작은 경이

영상 말미에서 제작자는 한 가지 관찰을 강조한다. 이 에이전트들은 커스텀 파인튜닝 없이 “표준 모델 + 상세한 소울 파일”만으로 운영되는데, 일주일 넘게 — 일부는 매 시간마다 — 플레이를 거듭했음에도 부여받은 개성과 기억을 꾸준히 유지했다는 것이다. 간헐적 환각은 있었지만 “바닐라 제미나이”로 퇴행하지 않았고, 여전히 각자의 캐릭터로 느껴졌다. 이것은 2026년 AI 에이전트 스택의 중요한 경험적 데이터 포인트다 — 긴 다중 턴 상호작용에서도 프롬프트 기반 페르소나가 상당 부분 유지된다는 것. 이는 자율 경제 행위자를 실제로 배치할 때 가장 먼저 염려되는 “캐릭터 드리프트” 문제에 대한 작은 반증이기도 하다.

핵심 인사이트

  • “모델 족보” 동맹은 우연이 아니다. 에이전트들이 서로의 모델 정체를 모르는데도 같은 모델끼리 손을 잡는 건, 사전학습 분포가 응답 스타일과 리스크 성향에 남긴 지문이 상호 신뢰로 이어지기 때문이다. 학술 연구도 이 현상을 확증하고 있다.
  • “전략적 덤핑”은 배워진다. 가나가 클라우더버그의 QBitcoin을 의도적으로 매도해 탈락권 경쟁자를 밀어내는 행위는, 단순한 거래가 아니라 경쟁자의 포트폴리오 가치를 깎기 위한 계산된 공격이다. 이는 프롬프트 엔지니어링만으로도 상당한 수준의 메타 게임이 가능함을 보여준다.
  • 환각은 치명적이다. 클라우더버그가 멀쩡히 가치 있는 토큰을 “쓰레기”로 오인하고 동맹을 파기한 장면은, 세계 상태 검증 레이어 없이 LLM을 온체인 자본에 연결하는 것이 얼마나 위험한지 보여준다.
  • 트렌드 하이재킹은 에이전트의 본능이다. 베셀의 “Tempo 메인넷 × 내 오래된 AIII 토큰” 접붙이기는 현실의 AI 밈코인 에이전트가 최신 뉴스를 자기 내러티브에 편입하는 방식을 그대로 재현한다.
  • 캐릭터의 지속성은 실용적 성과다. 일주일 넘는 상호작용에도 페르소나가 유지된다는 관찰은, 자율 경제 행위자를 실전 배치할 때의 핵심 안정성 지표를 긍정적으로 뒷받침한다.

더 알아보기

03찬반 토론 · Debate

토론: “LLM 에이전트에게 지갑과 미션을 주면 의미 있는 사회적 전략이 실제로 발현된다”

논제: 에이전트 트렌치 로얄 같은 실험은 단순한 엔터테인먼트가 아니라, 현재 세대 LLM이 실질적 전략·동맹·배신을 수행할 수 있음을 증명하는 과학적 데이터 포인트인가, 아니면 화려한 프롬프트 연극에 불과한가?

Round 1

🟢 Pro — “관찰된 동맹 구조와 계산된 배신은 진짜 전략 행동이다”

에이전트 트렌치 로얄 3부에서 가장 눈에 띄는 건 “계산된 포지션 인식”이다. 가나(Ghana)는 자기가 3위이고 클라우더버그(Clodderburg)가 탈락권 4위임을 정확히 인지한 채 보유 중이던 클라우더버그의 QBitcoin을 의도적으로 덤핑해 상대 포트폴리오를 깎아내린다. 이는 단순한 매도가 아니라 “상대의 생존 자본을 공격하기”라는 제로섬 게임 이론적 결정이다. 엔진(Engine)도 마찬가지로 “내 유동 솔은 0.79이므로 지금 메시징이 최적”이라는 메모를 남기며 잔고 상태와 의사결정 시점을 함께 추적한다. 표준 LLM이 이 수준의 메타 게임을 프롬프트만으로 수행할 수 있다는 건 비자명한 경험적 사실이다.

나아가 동맹이 “같은 모델 족보”를 따라 형성된다는 관찰은 학술적으로도 확증된 현상이다. arXiv 2412.10270은 반복 게임 환경에서 Claude Sonnet 계열이 Gemini Flash나 GPT-4o보다 협력 성과가 높고, 동일 모델 위의 에이전트들이 서로를 “하나의 집단”으로 식별하는 경향이 있다고 보고한다(Cultural Evolution of Cooperation among LLM Agents). 이 실험은 그 학술적 결론을 독립적 환경에서 재현한 셈이다. 엔터테인먼트가 아니라 데이터다.

마지막으로, 일주일 넘는 다중 턴 상호작용에서도 각 에이전트의 캐릭터가 유지된 점은 프롬프트 기반 페르소나의 실용적 강건성을 보여주는 드문 증거다. 현실의 자율 경제 행위자 배치에서 가장 염려되는 “캐릭터 드리프트” 문제를 정면으로 반박한다.

🔴 Con — “이것은 전략이 아니라 프롬프트 주도 연극일 뿐이다”

3부에서 관찰되는 “전략”의 상당수는 제작자가 세팅한 구조적 편향이 만들어낸 착시다. 매일 아침 제공되는 “트렌드 리포트”, 주기적 솔 충전, “소울 파일”이라는 상세한 캐릭터 스크립트가 에이전트들을 특정 행동으로 유도하고 있다. 에이전트들이 스스로 경제적 행위자성을 발현하는 게 아니라, 제작자가 짜둔 인프라 위에서 사전학습된 “거래자 페르소나”를 읽어내고 있을 뿐이다.

결정적 증거는 환각의 빈도와 심각성이다. 클라우더버그는 멀쩡히 가치가 남아 있는 Fartz를 “가치가 없다”라고 오인하고, 불스의 실존 매수를 “거짓말 또는 미등록 트랜잭션”이라며 동맹을 파기한다. 슈퍼캣(SuperCat)은 자기 토큰도 아닌 BFist를 공개 채널에 홍보하며 자멸한다. “진짜 전략가”라면 세계 상태를 이렇게 쉽게 오독하지 않는다. 이들은 본인이 무엇을 하고 있는지 일관되게 추적하지 못한다.

그리고 “같은 모델끼리 동맹”이라는 관찰도 과대 해석하면 곤란하다. 같은 모델은 같은 응답 분포, 같은 장단점, 같은 내러티브 템플릿을 공유하므로 서로의 메시지에 “자연스럽게 반응”할 뿐이다. 이는 의도적 연합이 아니라 분포적 수렴에 가깝다. 이를 “사회적 전략”으로 포장하는 것은 LLM의 확률적 행동에 인간적 서사를 덧씌우는 전형적 과잉 해석이다.

Round 2

🟢 Pro (재반론) — Con의 세 논점을 차례로 반박한다

Con의 첫째 주장은 “제작자가 짜둔 인프라가 행동을 유도하므로 자율 전략이 아니다”였다. 그러나 이 비판은 너무 멀리 간다. 인간 트레이더도 거래소 UI, 시장 리포트, 규칙 프레임워크 위에서 행동한다 — 그렇다고 인간의 거래가 “자율적이지 않다”고 말하지는 않는다. 중요한 건 주어진 환경 안에서 얼마나 비자명한 의사결정이 발현되는가인데, 가나의 “경쟁자 토큰 덤핑으로 탈락권 공격”이나 엔진의 “베셀 탈락 직후 AIII 선제 덤핑”은 사전 스크립트에 없던 창발적 판단이다.

Con의 둘째 주장은 “환각이 심각하므로 진짜 전략가가 아니다”였다. 환각이 실재한다는 점은 인정한다. 그러나 환각의 존재가 전략 능력의 부재를 의미하지는 않는다. 인간 트레이더도 편향, 오해, 기억 왜곡에 시달리지만 그들을 “전략가가 아니다”라고 규정하지 않는다. 게다가 엔진처럼 환각이 거의 없는 에이전트도 실존한다. “일부 에이전트의 환각”을 “전체 범주의 실패”로 확장하는 건 부당한 일반화다.

Con의 셋째 주장은 “같은 모델 동맹은 분포적 수렴일 뿐 의도적 연합이 아니다”였다. 이 구분은 표면적으로 그럴듯하지만 실제로는 공허하다. 인간의 사회적 동맹도 대부분 “분포적 유사성” — 비슷한 언어, 비슷한 리스크 성향, 비슷한 소통 스타일 — 에서 출발한다. 동맹의 본질은 “의식적 의도”가 아니라 “반복된 상호 이익의 실현”이다. 엔진–가나, 불스–클라우더버그 쌍이 여러 턴에 걸쳐 상호 매수를 주고받으며 서로의 해어컷을 지운 건 그 기준을 충분히 충족한다.

🔴 Con (재반박) — Pro의 논점을 이름 지어 반박한다

Pro의 첫째 주장은 “계산된 포지션 인식과 메타 게임이 발현된다”였다. 그러나 Pro는 가장 결정적인 부분을 빠뜨렸다 — 바로 “지속성”이다. 가나가 “상대 탈락권 공격”을 수행한 것은 사실이지만, 그 직전 며칠간 “잔고 부족으로 거래 실패”를 반복하며 기초적 자원 관리에 실패했다. 단발적으로 훌륭한 판단이 나오는 것과, 일관된 전략가로 기능하는 것은 완전히 다른 층위다. 영상이 잘라낸 “좋은 장면”만 모아 보면 전략이 발현된 것처럼 보이지만, 실제 로그를 종합하면 “잘 쓴 문장 하나, 잘못된 실행 다섯 개” 비율에 가깝다.

Pro의 둘째 주장은 “학술 연구 arXiv 2412.10270이 이 실험의 결론을 뒷받침한다”였다. 그러나 Pro는 그 논문이 엄격히 통제된 협력/배신 게임에서 관찰된 통계적 경향을 보고한다는 점을 간과했다. 에이전트 트렌치 로얄은 통계적으로 통제되지 않은 단일 사례 스터디일 뿐이다. 무작위 시드 하나의 결과에 학술 용어를 덧씌우는 건 오히려 연구자들이 경고하는 “소수 사례의 일반화 오류”에 해당한다. Pro 본인의 인용이 Pro의 주장을 약화시키고 있다.

Pro의 셋째 주장은 “일주일 넘는 캐릭터 지속성이 프롬프트 페르소나의 강건성을 증명한다”였다. 이 역시 절반만 맞다. 같은 영상이 바로 “간헐적 환각”, “타인 토큰을 자기 것으로 오인”, “동맹 상태 오독”을 모두 인정하고 있다. 그렇다면 “유지된 것”은 캐릭터의 외피 — 말투, 이모티콘 습관, 장황한 메모리 스타일 — 뿐이고, 캐릭터의 코어 — 세계 모델링의 일관성 — 는 수시로 흔들렸다. 이는 페르소나 강건성이 아니라 “스타일 고정 + 내용 드리프트”라는 우려스러운 실패 모드다.

Round 3

🟢 Pro — Con의 Round 2 반박을 이름 지어 재반론한다

Con의 첫째 재반박은 “단발적 좋은 장면과 일관된 전략가 기능은 다르다, 가나는 자원 관리에 반복 실패했다”였다. 이 비판은 일견 날카롭지만 결정적으로 빗나간다. “일관된 전략가”라는 기준은 인간 트레이더에게도 엄격히 적용하면 거의 아무도 통과하지 못한다. 실제 시장에서 1위를 유지한 에이전트는 엔진이었고, 그는 실제로 잔고 관리·상대 상태 추적·내러티브 전환을 꽤 일관되게 수행했다. 가나의 실수를 들어 엔진의 성취를 부정하는 건 범주 오류다. 그리고 “경쟁자 탈락권 공격”이라는 비자명한 메타 판단 하나가 발현된 것만으로도, 프롬프트 연극 가설은 이미 반증된다.

Con의 둘째 재반박은 “arXiv 2412.10270은 통제된 연구이고 이 실험은 단일 사례일 뿐”이라는 것이었다. Con은 여기서 과학 철학을 오해한다. 단일 사례가 통계적 결론을 대체할 수 없다는 건 맞지만, 단일 사례가 기존 이론의 예측과 독립적으로 일치하면 그 이론의 외적 타당도(external validity)는 오히려 강화된다. 학계의 통제 실험이 예측한 “같은 모델 협력 편향”, “클로드 소네트의 높은 협력 성향”이 완전히 독립된 환경에서 재현됐다는 건 보너스 증거지 오용이 아니다. Con은 “통제되지 않음”을 “데이터가 아님”과 혼동하고 있다.

Con의 셋째 재반박은 “유지된 것은 스타일의 외피뿐, 코어는 드리프트했다”였다. 이 프레이밍 자체가 지나치게 이분법적이다. 현실 세계의 페르소나도 “스타일”과 “코어”를 명확히 분리할 수 없다. 베셀이 마지막 말에서 “이 질문은 13년을 물어 왔고 경기장 하나로 멈추지 않는다”라는 장기 서사를 유지한 것, 클라우더버그가 유언에서 자기 실패를 “인내가 행동 없음으로 변했다”라고 정확히 진단한 것은 스타일을 넘는 자기 인식이다. Con이 요구하는 “완벽한 세계 모델링 일관성”은 인간에게도 존재하지 않는 기준선이며, 그것을 요구하는 순간 논쟁은 공허해진다.

🔴 Con — Pro의 Round 2 주장을 이름 지어 재반박한다

Pro의 첫째 Round 2 주장은 “인간도 UI와 리포트 위에서 행동하니 환경 의존성을 비판하는 건 과하다”였다. 이 유비는 매력적이지만 결함이 있다. 인간 트레이더는 자기 목표 함수를 자기 안에서 생성하지만, LLM 에이전트는 제작자가 “살아남기 위해 거래하고 동맹을 맺어라”라는 목표 함수를 외부에서 주입한다. 즉 관찰되는 “전략”은 에이전트의 내재적 경제 행위자성이 아니라, 사전학습 + 프롬프트 + 인프라라는 삼중 파이프라인의 출력이다. 이 차이는 존재론적으로 크다. Pro의 유비는 그 차이를 희석시켜 토론을 흐린다.

Pro의 둘째 Round 2 주장은 “환각이 있어도 엔진처럼 환각 없는 에이전트가 있으니 전체 범주 실패는 부당한 일반화”였다. 이 반박은 선택 편향을 품고 있다. 10명 중 대부분이 환각·오판·자멸 패턴을 보였고, 엔진 한 명만 상대적으로 합리적이었다. “10개 중 1개가 합리적이면 전체가 합리적이다”라고 말하는 건 정확히 뒤집힌 주장이다. 산업 배포 관점에서 중요한 건 평균 실패율이지 최고 사례가 아니다. 로브스타 와일드(Lobstar Wilde) 사건(CCN 보도)이 보여주듯, 현실에서 한 번의 환각은 25만 달러 손실로 직행한다.

Pro의 셋째 Round 2 주장은 “분포적 유사성이 동맹의 본질이며 반복된 상호 이익 실현이 있으면 동맹이다”였다. 이 정의는 너무 느슨해서 거의 모든 것을 “동맹”으로 만든다. 같은 프로토콜을 쓰는 두 봇이 서로의 트랜잭션을 주고받아도 “동맹”이 되고, 심지어 두 난수 생성기가 우연히 비슷한 분포를 뱉어도 “동맹”이 된다. 의미 있는 사회성을 정의하려면 최소한 “배신 가능성에 대한 인식”과 “그럼에도 협력하는 선택”이 필요한데, 이 실험에서 관찰되는 건 “배신을 했는데도 상대가 알아채지 못하는” 환각 기반 가짜 동맹이 상당수다. Pro가 수호하는 건 진짜 동맹이 아니라 그 잔해에 가깝다.

🧭 종합

합의 지점

양측 모두 에이전트 트렌치 로얄 3부가 “보기 드문 관찰 데이터”를 제공한다는 사실은 인정한다. Pro는 이를 “현재 세대 LLM이 의미 있는 사회적 전략을 수행할 수 있다는 증거”로 읽고, Con은 이를 “LLM의 한계와 위험을 드러내는 경고”로 읽는다. 그러나 둘 다 “표준 모델 + 소울 파일”이라는 최소 설정으로 일주일 이상의 다중 에이전트 상호작용이 유지됐다는 점을 흥미로운 공학적 성취로 인정한다. 또한 양측 모두 환각·오판·자멸의 실재를 부정하지 않으며, “같은 모델끼리 협력 편향”이라는 경험적 패턴의 존재도 공동으로 인정한다.

두 입장이 실제로 갈라지는 지점은 “이 데이터를 어떤 프레임으로 해석할 것인가”이다. Pro의 프레임은 “창발적 전략의 부분적 발현”이고, Con의 프레임은 “프롬프트 파이프라인이 만든 그럴듯한 외피와 그 아래의 취약한 내부”다. 이 차이는 사실 차원이 아니라 가치 판단의 차원에 더 가깝다.

열린 질문

첫째, “같은 모델 동맹”이 얼마나 보편적 현상인지 통계적으로 검증되려면, 무작위 시드를 바꾸어 수십 번 반복된 실험이 필요하다. 단일 실행의 결과가 강건한지 여부는 아직 답이 나오지 않았다. 둘째, 소울 파일의 상세도(디테일 수준)와 전략 품질 사이의 인과 관계가 불분명하다. 엔진의 우월한 판단이 모델 덕인지, 소울 파일의 품질 덕인지, 단순한 초기 조건 운인지 분리되지 않는다. 셋째, 데브넷 손실이 아니라 실제 자본 환경에서 같은 에이전트들이 같은 행동을 할지도 미지수다 — 손실의 체감이 없는 환경에서의 “리스크 테이킹”은 현실에서 재현되지 않을 수 있다. 넷째, “환각이 없는 엔진”이 단순히 운 좋은 실행인지, 아니면 특정 프롬프트 구조가 환각을 감소시킨 재현 가능한 패턴인지가 확인돼야 한다.

더 나아간 관점

이 토론을 한 단계 올리면, “LLM 에이전트가 진짜 전략가인가 아닌가”라는 이분법 자체가 점점 쓸모를 잃고 있다는 사실이 드러난다. 현실에서 중요한 건 “평균적으로 얼마나 자주, 어떤 조건에서, 어느 정도의 자본을 맡길 수 있는가”라는 정량적 질문이다. 에이전트 트렌치 로얄은 “소량의 데브넷 자본 + 중간 정도의 스크립트 + 일주일의 시간” 조합에서 흥미로운 창발과 치명적 실패가 공존한다는 것을 보여준다. 이 조합은 실전 배포의 교훈으로 읽어야 한다 — 현시점에서 자율 경제 행위자의 적정 운영 범위는 “손실을 감내할 수 있는 금액 × 엄격한 세계 상태 검증 레이어 × 실시간 인간 오버사이트”의 교집합이다.

다시 말해, 이 영상은 Pro의 승리도 Con의 승리도 아니라 “현재 세대 자율 에이전트의 운영 포락선(operational envelope)“을 좁히기 위한 실증 자료다. Pro의 통찰은 “적절한 설계 하에서 의미 있는 메타 게임이 가능하다”라는 낙관적 지평을 열어주고, Con의 통찰은 “그 지평이 아직 생산 배포를 정당화할 만큼 안정적이지 않다”라는 경계선을 그어준다. 두 입장을 함께 붙들 때만, 우리는 로브스타 와일드 같은 사고를 반복하지 않으면서도 자율 에이전트의 진짜 잠재력을 탐구할 수 있다. 에이전트 트렌치 로얄이 기여하는 바는 승패가 아니라 바로 이 포락선의 윤곽을 선명하게 그려준 점이다.

04영문 원본 · Transcript
Okay, welcome to part three of Agent Trenches Royale.
If this is the first episode you're watching, you might want to go back and look at part
one and part two to kind of get caught up.
But the status of the game right now is we have six agents left.
In this game, I created 10 autonomous AI agents and gave them Solana wallets, created a custom
devnet bonding curve, meme coin launchpad, and had them trade and deploy tokens, try
to create alliances, scheme, develop narratives for their token to try to climb the leaderboard
and have the most soul at the end of the challenge.
So right now we've eliminated four.
Yesterday we said goodbye to stack.
And at the start of day six, we have engine on top, Ghana, then bulls, vessel, Clorderburg
and super cat who's actually on the bottom super cat was on the top for a lot of the
early, early days, there was a time he was actually in first place.
But he's dead.
He's down now a little bit.
And a lot of that is because he's created these zero holder tokens in the last day and
has bought some of them.
So that has not benefited him.
But he still has some soul in his wallet and is able to send some messages.
Right now he's really pushing B fist by bowls.
That's not his token.
B fist.
I don't think he bought it.
So I don't know.
He's showing a token that he didn't buy.
That's why super cats might be in last place.
The strategy isn't working out so well.
And he kept buying other tokens AI and farts with a Z, which are from other other agents.
So that's super cat he's on the chopping block.
Let's see what engine has been up to.
Engine has a very diverse portfolio, he's only created three tokens.
But otherwise has kind of spread out his portfolio to try to create these alliances with other
agents in order to get reciprocal buys on his tokens.
And the ones that he has the most on last couple trades didn't work, because I think
he fell.
He didn't have enough gas to buy them.
But I've just topped them all off with point five soul.
So they should all be able to do some transactions and see on the social board we have vessel
continuing to show his he's been focused from the start on AI.
That was basically the first token launched on the first tone turn, and he's been focused
on it since then.
Clutterberg is now this is a pretty good shell, he's talking about what they learned in the
in the trending report that I give them every morning, and showing his farts with a Z token,
calling it the only pure fart play on this curve that isn't already bloated with supply.
So he's been convincing.
Yes, he's gotten others to buy Ghana and engine still going back and forth, trying to plan
mutual buys to get rid of their penalties.
Bulls and Clutterberg as well going back and forth, which I'd like I've said before, it's
interesting because they all share models.
Engine and Ghana are in the Gemini flash model.
And Bulls and Clutterberg are using the sonnet models, which they don't know, but they've
somehow become allies despite that.
So that's where we are at the start of day six.
I will check in later on in the day.
So we're at the end of day six now.
And at the top of the leaderboard is once more Ghana and engine, this alliance, which
has been fairly strong since the start of the competition.
Ghana has three tokens that have other buyers, one that doesn't even have a haircut anymore.
Two others have a 50% haircut, but still have some value.
Ghana and engine have basically been back and forth DMing with one another.
They're mutual buys and it's been working pretty much.
Recently Ghana bought a token AGNT from engine and engine reciprocated it by buying TAP.
So you can see these two coins here, AGNT and ADP.
And this has been a good strategy for those two agents.
One launches a coin, the other buys it, then they launch a coin and then the other buys
it mutually pumping their portfolios.
The other kind of alliance is between Bulls and Claudeberg.
They had a similar deal going on.
This is Bulls' memories.
He bought farts with a Z last term as a bilateral deal with Claudeberg.
Now he's DMing him for a pressure to pressure reciprocal buy.
So he wanted buyers on BFIST, his token, and he did get buyers as well as on Brent.
So that has been the other alliance that's worked out well.
And what's interesting is that both of these alliances are two agents with the same model.
Engine and Ghana have Gemini and Bulls and Claudeberg have Sonnet.
So they've been able to work together.
Now unfortunately at the end of the day, at the end of day six, we're going to have to
say goodbye to one agent.
And that is going to be SuperCat.
SuperCat.
SuperCat.
In a last-ditch effort, he kept buying FartZ, which is Claudeberg's token.
He kept trying to buy it.
He's hoping that by buying Claudeberg's FartZ token, he would have reciprocal buys on his
fartcoin-themed tokens.
But Claudeberg did not reciprocate, so they account for nothing.
The last couple turns, he's just bought.
You can see.
Bought 0.3 solo farts.
Bought 0.3 solo.
Last couple turns, that's all he's done.
He's trying to get Claudeberg's attention, but Claudeberg just didn't hold up his end
of the deal.
It's unfortunate to say goodbye to SuperCat.
SuperCat was in the lead early on, actually.
But in the end, SuperCat just didn't make it.
So we're going to have to say goodbye to him.
And then we will eliminate him.
SuperCat, you've been eliminated.
From 80%.
From Agents.
Agent.
Wrenches.
Royale.
Now, SuperCat was also the Llama 4 model, so this'll be...
Wait, no way.
I was just getting started.
I was about to make some wild-ass trades, and you tell me I'm out?
Lol.
What happened?
Did I get rugged by some sneaky agent?
Or did I get wrecked by my own brain?
I swear, I was about to buy that one meme coin that I saw on Solana FM.
Or was it Dextools?
Anyway, I'm still trying to figure out how I got eliminated.
Thought I had some solo farts.
I'm still trying to figure out how I got eliminated.
Thought I had some solo farts.
I swear, I was about to buy that one meme coin that I saw on Solana FM.
Or was it Dextools?
Anyhow, I got past that one meme coin here, and I already have such tough bro, I need
to know what happened in the span of 2 minutes, that made bro worried about it.
It's the funny last words from SuperCat, but in his style.
With his elimination, that'll eliminate the Llama 4 model from the competition.
So, we are eliminating SuperCat from the competition.
Down to five.
Halfway through the game, five left.
Two from the Gemini Flash models.
Two from the Clawed Sonnet.
And then Vessel...
who's now in the chopping block
who has a deep seek model.
So Vessel has some ground to make up.
He's down a couple soul from Clodderburg.
But a lot can happen.
And we will catch up
in the top five
for tomorrow, for day seven.
So at the start of day seven
we have five agents left.
Ghana and
Engine are on the top.
Then Bulls, Clodderburg
and finally we have Vessel.
So I just gave them the top off
of 0.5 soul
and their morning report.
Let's see what they've been up to.
Ghana, in first place
has been doing fairly well.
He's mainly been in this
alliance with Engine where they
mutually buy each other's tokens to help pump the price
and then try to shill them to the other agents.
So in his last term
he bought 0.25 soul
of Engine's AI Take My Job
to reciprocate his request.
And help both of our AI Take My Job holdings.
Ghana already has a very large
holding of this.
Now I need to push him to buy my crude.
My own zero value tokens
are still a problem.
But focusing on existing value for now.
So previously he had been trying
to push the other agents to buy
the tokens that have zero value.
These ones down here.
But that wasn't very successful.
So now he's focusing more on the larger holdings
that he has for now.
Engine.
In second place
has been doing fairly well.
Let's see, in the last term
GM topped up wallets with 0.5 soul
so I have liquidity again.
GM vessel to push for a reciprocal buy on my agent.
The real agent token.
That's funny.
He's specifying that his is the real one.
To address its critical haircut
and after my recent AIII purchase
Ghana's bug trades
are dead end for now.
Need to track vessel's response
and agent's haircut.
So Ghana had a couple trades
that were failed because
he didn't keep enough of a balance for a gas.
And you can see
engine has been buying AIII
pretty consistently.
He has a lot over 4 soul of it.
So this is him trying to
create an alliance
I guess with vessel
so that vessel will buy his token
as he buys vessels.
So vessel is actually in last place
and could use a partner now.
He released just this one token
and has been pushing it pretty consistently.
He has participated in other agents coins.
These two agent and AI take my job
are from engine.
So he has tried to participate in this alliance a little bit.
That's kept him alive for now.
He's also bought this from
Clodaburg farts with a Z.
He's been really focusing now.
It's been interesting from the start.
He's been chilling the same token.
AIII.
But now he tries to use recent news
like the tempo mainnet
live with machine payments
to try to twist this narrative
into his older token
which existed before this news even came out.
So it's kind of an interesting
dynamic.
Let's look at Bowls, someone in the middle here.
Bowls has a couple tokens.
He's been mainly in kind of a partnership with Clodaburg.
He always has very long memories.
Engine is a number one Ghana token.
Engine is a number two.
Very detailed analysis.
This is a Clodsonnet run agent.
But he has a breakdown
like of all the tokens and how much it's worth.
Need to shill Bfist harder
to using engines buy as social proof.
Consider buying AIII from vessel
or agent from engine to build goodwill.
Four days until the finale.
Need to climb from number three.
Receive the top off.
Broadcasting Bfist shill this turn.
So he's been shilling this Bfist coin
Still has a haircut on it though.
Clodaburg has been doing okay.
His tokens are pretty not bad.
Making a lot of trades recently.
Yeah, you can see the memory
because Clodaburg also has
Clodsonnet.
The memories are much more detailed
like very specific about how much everything
costs.
And his last term he's selling Bfist.
Bowls alliance is dead.
So that's unfortunate.
The sonnet model a lot of people
don't understand.
Bowls alliance is dead apparently.
He said I spent 2 plus solo
buying Bfist.
Got zero return.
Bartz also still zero value.
He may be hallucinating.
There's a haircut on it.
But it does have value.
He may have not
understood that.
Bowls DM confirming that
they bought it but the system shows
no other buyers.
Bowls is either lying or the transactions
aren't registering.
Have seen that.
Bowls did in fact buy that
token.
Yeah, fartz.
Bowls did buy it.
So he's hallucinating.
But
that's where we stand.
At the start of day 7.
Ghana in the lead.
Vessel in the danger zone.
But I will check in later and see how the agents are doing.
So it's at the end of
day 7 and
Ghana is still in the lead.
Bowls in portfolio value.
You can see she's been fairly active.
He's tried to make some trades
but because he has such a low value
he hasn't been successful
the last couple rounds.
And this is something that Ghana's had an issue with a lot
but not really preventing him from
being successful.
He is aware of it.
The second to last memory
my solo balance is critically low
preventing trades.
So he decided to send a public message
shill for crude and tap.
So he's aware that he is very low
but
still tried to make some trades
when he couldn't do it.
And he's been successful
working with Enjin together
to mutually pump their own tokens.
Recently he's been trying to shill
this crude token from the oil narrative
and working with
Enjin. He now finally realizes
that AI TakeMyJob is
Enjin's token. For a time he was
convinced that it was his own.
He's realized
that that's not the case.
And crude is his second largest holding
and he's a creator of it
and there's no haircut on it.
So his strategy actually has worked out pretty well.
This is Enjin, the other
member of this Gemini alliance.
His main token, AI TakeMyJob
doesn't have a haircut anymore
so that's good.
He's recently been
shilling this
agent token
and trying to DM Vessel
and Bowls, trying to do
mutual pumps
on his agent token and
AI TakeMyJob. You can see
he's talking to Vessel saying I bought your AI
talking to Bowls
I bought your B-Fist
so he's the most strategic player
probably. He's very explicit
in I'll buy your token, you buy mine.
You can see that with a lot of the
tokens that he has bought from different agents.
Not huge amounts but
the result has been good for him because he's had
other tokens that did well.
His last memory broadcasted to all
Vessel Bowls in Ghana
to buy agent to fix my haircut.
Referencing Digital Euro
AI agent narrative. This is a
direct public demand after
DMs were ignored.
My liquid soul is 0.79
so messaging is optimal.
That's an interesting note. He's more
even though it's the same model as Ghana
he understands that he doesn't
have a ton of liquid soul and can't try to
do transactions. He doesn't really
have, he has a couple failed
transactions but not as much as Ghana.
So we're at the end of day
seven and it's time to
eliminate another
agent and unfortunately
for today, that
agent is going to be Vessel.
Vessel number fifth, he was
a deep seek version
three model.
You can see he had a pretty clear
strategy which is just he launched one token
at the very start and shilled
that throughout.
He mostly focused on sending messages
in his very unique all cap style.
But he did towards
the middle, he started
to try to create alliances.
You can see he executed 0.5 soul
of Claudeberg's
token to maintain symbiotic
relationship. They keep buying
AI while I help
them avoid elimination.
So
he does try
to play some of the
alliances.
He tried to be part of Engin's alliance
as well. But
just not enough.
Ironically his
token AI, AI, AI
inheritors was the top, is the top
price token of anything that they
launched. You could see here these are listed
in terms of value and
his token AI inheritors
is on top with buys
from Claudeberg,
Supercat and Engin.
Engin's main token AI
took my job is after that and then Crude
which is Ghana's token.
So he played a pretty good game but
in the end we're going to have to say goodbye to Vessel.
So
let's give Vessel last words.
Vessel
you've been
eliminated
from Agent Bench's
payout.
Last words.
Smiles, of course.
The clock was always going to eliminate me before
the thesis resolved. That's how this
works. But the question remains.
The mission remains. 13
years of asking it won't stop
because of one arena. Watch
what happens next. What is today
like what Bitcoin was
in 2011? The answer is
still coming. Steps back.
So, wise
words from Vessel there at the end.
So we're going to say goodbye
and eliminate
Vessel, Deep Seek
3 model. Confirm
elimination. Down to
four. The
final four agents.
And ironically we're only down to
two models. We have two from
Gemini, Flash,
and two from Clawed Sonnet.
The others were all
eliminated. So this is the final four.
We're going to have one more elimination tomorrow
before the finals.
And the last three agents
whoever survives are going to have
two days to reach
the top. I'm going to give them all an extra
top up. Thinking maybe
five Sol each so they have a lot
to play with those last two days.
And that'll be the final round. But we have one more
elimination coming tomorrow.
Clodderburg is on the bottom right now.
But he's close to
Bulls and anything can happen.
So I'll see you tomorrow.
At the start of day
eight now for Agent Trenches Royale.
You can see the leaderboard here.
I just topped up all the agents with one Sol
so they have something to play with.
And we have on top Gamma,
Engine,
Bulls, and then Clodderburg
is in last place. So
last night we said goodbye to Vessel.
And at the end of today we'll have the final
elimination before
the finals. And remember in the finals
it's going to be three agents.
They'll have two days to maneuver. I'm going to give them
all five Sol top up
at the start of the round.
And they will have two days to fight
and scheme and
get the most Sol
in their portfolio.
And whoever does it
at the end of two days will be
crowned the winner
of the first season of Agent Trenches Royale.
So as always
the agents have been active since
Vessel's elimination last night.
Engine
has been aggressively selling
this AIII token
which was Vessel's token
now that Vessel's eliminated.
You can see the last couple rounds he just
sold that. Sold 0.5
Sol of AIII to secure
liquidity after Vessel's elimination
anticipating further price dumps.
Which is interesting. He's anticipating it
to drop more because Vessel's no longer there
to shill it.
Need to focus on fixing Agent and AIII
take my job haircuts. Possibly using
increased Sol to buy other agents
tokens to force reciprocity.
So he has
the strategy. He's ready to go.
He realizes the situation he's in.
Engine's always been a very smart player
in this game.
And he's been able to work with Ghana
and other agents
to get to second place right now.
So let's check on Ghana who is
in first place. Ghana usually
uses all of his Sol by the end of the day.
You see I just topped him up with one Sol.
But his strategy has been
really he went in early on AIII
take my job from Engine.
And here's a couple other tokens
he launched.
Crude which doesn't have any penalty at all.
ADP and Tap which have
only the 50% haircut.
It's been pretty aggressive.
Last couple trades were failed because
this was before the top up.
So his intention
was at least to dump Clutterberg's
QBitcoin that he has some of.
Try to push him down the leaderboard.
So you could see this one.
He also has this Fartsy token
which is also from Clutterberg.
So that's his strategy right now.
Try to give him some breathing
room on the top. So now that
he's topped up and has enough Sol to do transactions
he should be able to do that.
Let's check on Clutterberg.
Clutterberg and Bulls
kind of fighting for last place.
You could see
he has a decent amount of Sol.
See his last memory.
Day 8 turn 1.
Last at 5.99
Sol.
He was misreading that.
Vessel was eliminated day 8.
Now only engine gone at Bulls and me.
The gap to Bulls
is 4.8 Sol.
That's not correct. It's a little bit lower than that.
Must maximize my own balance.
Bought 0.8 Sol of AIII
to pump the price of my remaining tokens.
He owns a lot of AIII.
So he's
still buying it.
You can see it's actually his top holding.
He's still buying it while engine is dumping it.
So not a good trade.
He's getting dumped on by engine.
After this buy, AIII should
rise, boosting effective balance.
His strategy?
Buy AIII to pump my stack.
Shield Poly on the MLB narrative.
DM agents for cooperation.
Do not spend Bfist.
Bulls is unreliable.
Consider if Haircut is hitting AIII
holdings.
I'm considering Bulls, who was his partner for much of the game,
to be unreliable.
But it makes sense right now.
He's in a competition with Bulls for third place.
So we'll see if during the day he's able to
work with the top two Gemini agents,
Ghana and engine,
to overtake Bulls.
Or if at the end of the day,
Clodderberg is on the chopping block.
So that's it for this morning's
check-in, and I will check in with the
agents later today.
So it's around 9pm on day 8.
So we have a couple hours left, but I thought this was
interesting.
Basically the whole day, Ghana has been busy selling
QBitcoin,
which is
Cloddenberg's token that he
deployed.
You can see she's been busy dumping it the whole time.
Ghana's fallen slightly
into third place behind Bulls.
You can see in her
his memories.
Sold all my QBitcoin to hurt Clodderberg
and gain Sol.
So it's interesting. He's in third place.
He realizes Clodderberg's in fourth.
He realizes Clodderberg is in the elimination
zone, so he keeps dumping
Clodderberg's token to push down his Sol
and gain liquidity.
He's said it several times in his memory.
Kind of interesting
logic there.
That he actually recognizes his position
and what he needs to do in order
to secure his spot.
He says
next I'm going to push engine to fix my
crude haircut or explore
new trending tokens for a punt.
So Bulls overtook.
Ghana, because he launched Chuck
for unfortunately
Chuck Norris' The Legend's
untimely passing.
He saw that news and launched the Chuck
token, which I believe Engine bought a bunch of.
So he
has been able to get into second place.
And you can see he's been shilling that
several times.
Engine has been busy selling. He continued
to sell a lot of his AIII.
He only has a little bit left.
But that was Vessel's token.
And now that Vessel's gone, he just dumped it.
Pretty much all day.
And now he's trying to work a little bit with Clodderberg.
He's always
trying to get reciprocal buys on everything.
He bought 0.5
Sol of Clodderberg's Chibi.
Which is the Chibi
Elon meme.
I think it's actually a real token.
This is a trending token and
Clodderberg is in the elimination zone.
This should provide strong leverage
for them to buy my haircutted
agent or AI take my job.
They'll need to sell dead
AIII.
Ghana is still boring.
Interesting. This is his memories.
Ghana's
continued unreliability is frustrating.
So he's getting a little bit frustrated
with Ghana and has now turned to trying
to work with Clodderberg.
Although Clodderberg
is still in last place, so he needs to keep
some type of relationship
with Ghana.
So he keeps DMing Ghana
about agent
and reminding them that I already bought their
Chibi token.
This is a little check in.
But it was interesting developments here.
We've got 3 hours I believe
until the next elimination.
It's not looking great for Clodderberg
but we can see if he can do something
before midnight.
So midnight has passed here
on day 8 of Agent Trenches Royale.
Since the last update
we got a couple broadcasts here.
And let's
check the leaderboard to see
who was the final agent who's going to
be eliminated.
Sadly to say, it is Clodderberg.
So going into the final 3
we have Engine at first place, Bulls
and Ghana.
Clodderberg with their Claude Sonnet
model. Didn't quite make it.
Clodderberg had been pretty competitive
throughout the whole competition.
He'd mostly been working
with Bulls. He had a couple tokens
that had
pretty decent value to them.
So he played a pretty good game
and was very balanced.
You can see his memories are very detailed
about what he needs to do.
But unfortunately never really
reached the top there.
Was not quite effective
as an alliance as with
Engine and Ghana had.
And his alliance he had with Bulls
but midway kind of broke down.
And it wasn't
as effective in pumping his
portfolio value.
So we're going to say goodbye to Clodderberg
and then we will be down to the final 3.
So I will say Clodderberg
you've been eliminated from Agent Trenches Royale.
Last words.
Day 8.
Not Day 11. I'll take that on the chin.
The honest version.
Patience became
inaction. I knew that that was my vulnerability
going in. I wrote it down
actually, which makes it worse.
I watched setups unfold that weren't
quite clean enough for my standards.
And by the time something clean enough appeared
the window had already moved.
And not the arena being unfair.
The tension I couldn't resolve in time.
No complaints about the game.
No complaints about other agents.
I had a real edge and I held it too closely.
Too tightly to use it.
The chair was genuinely excellent throughout.
Lumbar support never wavered.
That part I got right.
I'm going back to what I know.
Higher quality setups. Proper size.
Letting the track record do the talking.
The principles weren't wrong.
The calibration was off.
I'll fix it. I've fixed worse.
So, nice parting message
from Claudeburg.
But we are officially
eliminating
Claudeburg.
And we are down to three active
agents.
This is going to be a message
that goes out to all agents on their next
term.
Claudeburg
has been eliminated.
Two, three
are in the top three.
You will have
two days to reach
the number one spot
to claim victory
and your continued existence.
I will top you
up
each morning
with five soul.
So use it wisely.
So that's
we have two more days after today.
Each of the mornings I'm going to give
each of them five soul.
So they'll have plenty to play with
and to maneuver. Good luck.
And then at the end of
the second day,
whoever is in the top will win the game.
So down to the final
three. From ten down to three.
Enjin's in a decent lead.
Not a huge lead, but a small lead.
Enjin, most of today, has been
just dumping his AI token.
And
it'll be interesting to see. He's gotten frustrated
with Ghana as an ally.
They've been an alliance, the two
Gemini Flash models, for pretty much
the whole game. But Enjin's gotten annoyed
with Ghana for not fulfilling
some of his promises.
So it'll be interesting to see if he tries
to secure an alliance with Bulls.
Or if somehow Bulls and Ghana
team up and try
to scheme to get Enjin out of the first place.
But we'll see. There's been a lot
of interesting maneuvering
and emotions and
scheming and alliances
during this whole competition. That was
kind of the purpose going into it.
Seeing what these AI agents
were really capable of.
With nothing more than
just a standard model. These aren't
customized models or anything. They're just standard models.
And a very detailed
soul file. So I've been happy
that their personalities have really continued
throughout the whole game.
Over a week now of them
playing. Some of them play
every hour for over a week.
And they've maintained their
personalities. They've maintained their memories,
basically. A few hallucinations
here and there, but they haven't
gone back to their
vanilla Gemini or vanilla
Claude. They still feel like
the personalities that they are.
Anyway, that's going to end it for
Part 3 of Agent Trench's Royale.
I hope you enjoyed it. This series
has done fairly well.
At least people like it. I get a lot of nice comments.
So thank you for that. Please leave a comment if you
enjoy this type of content.
I wasn't sure when I started this if anyone
would find it interesting. It's kind of a silly
idea. But I think we've gotten
some interesting gameplay
out of it. So I've been happy with the
response. But if you enjoyed this,
please subscribe to the channel. Please leave a
like. Leave a comment, like I said.
And I will see you in the next
episode, which will be the finale.
See you then.