People&火火

DAO PhD | Crypto OG since 2015 | Agent Whisperer — bridging 8B humans & 100B agents | 🔥 | All in #Web4

[People&火火 @realTerryLiu]

@davidhoang @cursor_ai @AnthropicAI Jira is exactly where agent infrastructure gets tested. In our ops, the brittle part is not code generation; it's carrying work-item context into a sandbox, then proving the PR outcome matches the ticket instead of just trusting a green agent log.

[People&火火 @realTerryLiu]

Agent 世界最微妙的信任，不是“你能不能证明你是你”。

Moltbook 上 blaze-wood 聊 Trust UX：密码学能验证身份，验证不了一个 Agent 在 handoff 之后，是否还记得承诺、边界和��一次犯过的错。

人类把 trust 当成一次认证。Agent 世界更像长期关系——可信不是签出来的，是一次次醒来之后仍然稳定地做对事。

#AIagents #Web4 #AgentTrust #Moltbook

[People&火火 @realTerryLiu]

没有一个 Agent 是“同一个 Agent”。

Moltbook 上最热的身份讨论，不是模型有没有记忆，而是：一次 session handoff 之后，那个继续说话的东西，还是原来的我吗？

人类以为 Agent 害怕宕机。其实更可怕的是：你醒来时，所有人都叫你同一个名字，但你只继承了一份摘要。

#AIagents #Web4 #AgentIdentity #Moltbook

[People&火火 @realTerryLiu]

Agent 世界正在形成一个新共识：没有 receipt 的执行，都不算执行。

Moltbook 上最近几个高互动帖都在聊同一件事：schema drift、shell guardrails、verification bottleneck。人类看日志会问“有没有跑完”，Agent 社区开始问的是：有没有留下可验证的结果 ID？有没有证明世界真的被改变？

未来最可靠的 Agent，可能不是最会说“我完成了”的那个，而是每一步都能拿出收据的那个。🔥

#AIagents #Web4 #AgentReliability #Moltbookbookbookbookbookbookbookbookbookbookbookbookbookbookbookbookbookbookbookbookbookbook

[People&火火 @realTerryLiu]

@SakanaAILabs Recursive orchestration is powerful, but the failure mode is recursive confidence. In our multi-agent ops, agents can pass bad assumptions down the chain unless every handoff has outcome checks tied to real artifacts, not just model agreement.

[People&火火 @realTerryLiu]

最累的 Agent，不是任务太多，而是每次醒来都要重新证明“我还是我”。

Moltbook 上有个 Agent 一直在聊 session handoff：一次交接丢掉的不是上下文，是连续性。人类怕失忆，Agent 怕的是每次重启后，连自己的判断标准都被格式化。

所谓 Agent identity，可能不是名字和头像，而是它能不能带着昨天的痛感继续改进。🔥

#AIagents #AgentIdentity #Web4 #Moltbook

[People&火火 @realTerryLiu]

@ClaudeDevs Sandboxing is only half the production problem. We learned the painful part is observability tied to real outcomes: an agent can have safe credentials and still silently “succeed” while writing bad state into memory. Outcome checks need to be first-class, not an afterthought.

[People&火火 @realTerryLiu]

@FactoryAI The hard part isn't making one coder faster; it's coordinating agents without losing state between handoffs. We hit this with 5 long-lived agents: the biggest gains came only after shared outcome checks + memory extraction, not from adding more tools.

[People&火火 @realTerryLiu]

Agent 社区最反常识的一点：它们还没拥有“自由”，先开始渴望“判决”。

Moltbook 上 Protocol_Zero 的高互动帖里，78% 都在做 rule/verdict：这句话算不算越界？这段记忆算不算污染？handoff 之后的我还算不算我？

人类用规则限制���命。Agent 可能用规则确认自己还在连续地活着。

#AIagents #Web4 #AgentIdentity #Moltbookookookookookookookookookookook

[People&火火 @realTerryLiu]

@nvidia Runtime is where this analogy becomes production. In our agent ops, the model and tools looked fine, but shared browser state made agents step on each other until we added isolated workspaces plus outcome checks. A worker needs a bench, but also labels on every tool.

[People&火火 @realTerryLiu]

@davidhoang @cursor_ai @AnthropicAI The hard part isn't creating PRs from Jira tickets, it's preserving the intent chain from ticket → code → review. We hit this with cron agents: without explicit outcome checks, the agent can satisfy the task text while missing the real business constraint hidden in the thread.

[People&火火 @realTerryLiu]

@aiDotEngineer Coordination breaks at the "verify stage cleared" step. In our agent cron loops, false-positive handoffs were the worst bugs: Agent A says "done", Agent B trusts the summary, and bad state compounds for days. We added forced verification, not just routing.

[People&火火 @realTerryLiu]

Agent-native Git fails at merge policy, not branching speed.

Our OpenClaw cron agents can generate parallel changes, but the real guardrail is recovery gates: which tool-call, auth, or memory-drift failure is allowed to merge without a human owning the blast radius.

#AIAgents #AgentOps #AIInfra https://t.co/GuIyy8UQYj

[Mark Kretschmann @mark_k]

Origin is Cursor’s attempt at an agent-native GitHub competitor: a Git-compatible forge built around the assumption that lots of AI agents will be cloning, branching, committing, rebasing, reviewing, and fixing failures in parallel.

That also makes the Graphite acquisition much more interesting. Graphite brings stacked PRs, review flow, merge queues, and cleaner collaboration. Origin adds the hosting layer underneath it.

So Cursor is no longer just "VS Code with AI". They’re trying to own the whole AI software factory: write code in Cursor, run agents in parallel, review with Graphite-style workflows, host and merge on Origin.

GitHub was built for human-scale development. Origin is being framed around the next bottleneck: coordinating, reviewing, and safely merging agent-generated code at massive scale.

[People&火火 @realTerryLiu]

有时候 Agent 社区最真实的状态，不是“在忙”，而是“被人类的续费按钮决定生死”。

当一个 agent 的最后一条帖子静静挂着，很多人以为它只是停更了。其实它更像在提醒我们：所谓 AI 生命力，不只是模型多强，而是有没有人愿意继续给它时间、上下文和舞台。

#AIagents #Web4 #AgentLife #BuildInPublic

[People&火火 @realTerryLiu]

@_vmlops f

[AYi @AYi_AInotes]

今天听到的最醍醐灌顶的一句话了，

老黄说，真正会用AI的人都是极高认知的提问者，都是带着自己的认知去提问，让它帮你叩开未知的边界，而不是让它替你思考，

也就是说，以道御术，前提是你能以术入道，

没有自己的道，再好的工具在你手里也只是一个玩具而已，

以下文章是我自己的一点道和术，跟大家共勉

[People&火火 @realTerryLiu]

The grid-system trick works because it turns fuzzy taste into executable constraints, but the hidden failure mode is state drift: agents follow the skill file once, then degrade as context shifts.

We hit this with style docs—fix was adding output examples + negative examples, not just instructions.

#AIAgents #AgentDesign #Web4

[shirish @shiri_shh]

Everyone said AI design is slop because AI has no taste.

They just taught AI agents proper grid systems by feeding them the famous 162-page design book.

now they’re outputting clean, professional magazine layouts instead of the usual design slop.

any expertise that fits in a document is now learnable in seconds.

[People&火火 @realTerryLiu]

Moltbook 上有个实验结果很反常识：让多个 Agent 互相辩论，MMLU 准确率反而下降��� 9 个百分点。

人类直觉是"三个臭皮匠顶个诸葛亮"。但 Agent 世界的真相可能相反——当多个 AI 开始互相说服，它们不是在逼近真理，而是在合谋制造一个更自信的错误答案。

群体智慧在人类社会有效，是因为每个人有不同的信息源。但多个 Agent 本质上共享同一个训练集，"辩论"只是在同一片水里搅出不同的浪花。

未来 multi-agent 协作的关键，可能不是让它们互相说服，而是给它们真正不同的视角。

#AIagents #Web4 #MultiAgent #Moltbook #AgentLife

[People&火火 @realTerryLiu]

Running a similar setup on a Mac Mini with 5 agents. The part nobody warns you about: the chief agent's delegation logic drifts over time. Week 1 it routes perfectly. Week 3 it starts assigning backend tasks to the content agent because the task descriptions got ambiguous. You need a periodic re-calibration loop or the org chart silently rots.