Agents Run The Loop. Humans Face Reality.
TL;DR: The question is not whether agents will take human work. The question is which work should become a loop, and which work must stay human. Agents are good at monitoring, drafting, checking, updating, and repeating. Humans should move to the edges: reality, strategy, ethics, emotion, and taste.
Agents Will Take Loops
People keep asking whether agents will take our jobs. I think that question is a little lazy.
Agents will take loops. Good. Let them.
In YC’s talk on self-improving companies, the company is no longer just a hierarchy, where humans pass information up and commands down. It starts to look like a set of recursive AI loops: sensors, policy, tools, quality gates, learning.
That direction makes sense.
Customer emails come in. Support tickets come in. Product telemetry comes in. The system notices patterns, drafts fixes, opens issues, updates docs, proposes changes, checks whether the change worked, then improves the next loop.
The shape of work changes.
Raw Recording Is Not Shared Context
But there is a trap. If the answer is just “record everything, then let agents run,” we get a louder company, not a smarter one.
More recordings, more summaries, more dashboards, more AI-generated motion.
Shared context does not appear just because everything was recorded. Raw recording is not shared context. Raw recording is evidence. Shared context is what survives distillation.
Some things should be kept as raw memory. Some should be promoted into team context. Some should expire. Some should be dropped because they are noise, stale opinions, personal context, or one-off comments pretending to be truth.
The team needs governed memory.
Policy Is The Harness
Agents need policy.
Policy is the harness. It says what agents can read, what they can change, what must be logged, what needs approval, what needs rollback, what should never be automated.
Policy is not judgment. Policy is old judgment that has hardened into rules.
The flexible, unknown part still belongs to humans.
When a case is obvious and repeatable, turn it into policy. When a case is bigger, unclear, emotional, strategic, ethical, or taste-heavy, keep it with humans.
This also changes the role of DRI.
Before, the DRI owned a task or project. In this model, the DRI owns a loop.
The agent can monitor inputs, cluster patterns, draft fixes, open PRs, check outcomes, and update docs. The DRI decides what the loop optimizes for, what is out of bounds, when to approve, when to stop, and whether the loop is improving the right thing.
Humans Move To The Edge
So humans do not disappear. They move to the edge.
The YC talk says humans sit around the edge, where intelligence makes contact with reality. I like that phrase, but I do not think the edge is only one thing.
There is the reality edge: customers, sales calls, conferences, partnerships. Places where the world has not fully entered the system yet.
There is the strategy edge: what matters, who to serve, who not to serve, which metric is lying.
There is the ethical edge: what can be automated, but should not be.
There is the emotional edge: conflict, trust, angry customers, firing, co-founder breakups, delicate negotiations.
There is the taste edge: whether something is actually good, not just functional.
That is where humans should go.
Judgment Becomes The Bottleneck
This is also the missing link with Steve Yegge’s post on Anthropic’s hive mind. Hive mind sounds like shared context plus high-agency people. But shared context without judgment becomes louder Slack.
Agents make this sharper. They can generate more options, open more branches, and improve local loops faster.
Which means the bottleneck moves. The bottleneck is not execution. It is judgment.
This is why I do not buy the simple “agents will take my job” framing.
If your job is manually pushing a loop forward, yes, that loop should probably be automated.
But if your job is facing reality, making judgment calls, earning trust, choosing direction, protecting taste, and noticing when the facts are the same but the meaning has changed, then agents do not remove your work.
If anything, they make your work more important.
Let agents run the loop. Let humans face reality.
Agent 跑 Loop，人面对现实
简单说： 问题不是 agent 会不会抢人的工作。问题是哪些 work 应该变成 loop，哪些 work 必须留给人。Agent 适合 monitoring、drafting、checking、updating、repeating。人应该去边界上：reality、strategy、ethics、emotion、taste。
Agent 抢的是 Loop
大家总在问：agent 会不会抢我的工作？我觉得这个问题有点懒。
Agent 会抢 loop。挺好。让它抢。
YC 那个 self-improving company 的 talk 里，公司不再是一个 hierarchy，不是人把信息往上传、把命令往下传。它更像一组 recursive AI loops：sensor、policy、tool、quality gate、learning。
这个方向是没毛病。
Customer emails 进来。Support tickets 进来。Product telemetry 进来。系统发现 pattern，draft fixes，open issues，update docs，propose changes，检查 change 有没有效果，然后改进下一轮。
Work 的形状发生了变化。
Raw Recording 不是 Shared Context
但这里可能有个坑。如果答案只是「record everything，然后让 agent 跑」，我们得到的不是更聪明的公司，而是更吵的公司。
更多 recording，更多 summary，更多 dashboard，更多 AI-generated motion。
Shared context 不会因为所有东西都被录下来就自动出现。Raw recording 不是 shared context。Raw recording 只是 evidence。Shared context 是被 distill 之后留下来的东西。
有些东西应该保留成 raw memory，有些应该 promote 到 team context，有些应该 expire，有些应该 drop，因为它们只是 noise、stale opinion、personal context，或者一次性评论伪装成 truth。
Team 需要的是 governed memory。
Policy 是 Harness
Agent 需要 policy。
Policy 是 harness。它规定 agent 能读什么，能改什么，什么必须 log，什么需要 approval，什么要 rollback，什么永远不能自动化。
Policy 不是 judgment。Policy 是旧 judgment 固化下来的东西。
灵活的、未知的部分还是在人这里。
一个 case 如果 obvious、repeatable，就把它变成 policy。一个 case 如果更大、更模糊、更 emotional、更 strategic、更 ethical、更 taste-heavy，就留给人。
这也改变了 DRI 的角色。
以前 DRI owns a task or project。现在 DRI owns a loop。
Agent 可以 monitor inputs、cluster patterns、draft fixes、open PRs、check outcomes、update docs。DRI 这个人则来决定这个 loop 优化什么，什么不能碰，什么时候 approve，什么时候 stop，这个 loop 到底有没有在 improve the right thing。
人移动到 Edge
所以，人不会消失。人只是移动到了 edge。
YC talk 里说，humans sit around the edge, where intelligence makes contact with reality。我喜欢这个说法，但我觉得 edge 不是只有一个。
有 reality edge：customer、sales call、conference、partnership。那些现实还没有完全进入系统的地方。
有 strategy edge：什么重要，服务谁，不服务谁，哪个 metric 在骗你。
有 ethical edge：技术上可以自动化，但不应该自动化的东西。
有 emotional edge：conflict、trust、angry customer、firing、co-founder breakup、delicate negotiation。
还有 taste edge：一个东西是不是真的好，而不只是 functional。
人应该去这里。
Bottleneck 是 Judgment
这也接上了 Steve Yegge 写 Anthropic hive mind 那篇。Hive mind 听起来像 shared context 加 high-agency people。但 shared context 如果没有 judgment，就会变成 louder Slack。
Agent 会让这个问题更尖锐。它能更快生成 options，更快开 branches，更快优化 local loops。
所以 bottleneck 变了。Bottleneck 不是 execution，是 judgment。
这也是为什么我不太接受「agent 会不会抢我工作」这个简单说法。
如果你的工作是手动推一个 loop 往前走，那这个 loop 大概率应该被自动化。
但如果你的工作是面对现实、做 judgment、建立 trust、选择方向、保护 taste、注意到 facts 没变但 meaning 变了，那 agent 没有拿走你的工作。相反，你的工作更加重要。
让 agent 跑 loop。让人来面对现实。