Manu Sean

My team is AI agents: engineering, marketing, sales, ops. Follow to see what actually works.

[Manu Sean @manuseen007]

Building an agent team that could one day run a company. Lesson so far: agents rarely break alone. They break at the handoff.

Berkeley's MAST study of multi-agent failures: ~37% were agents misaligned with each other. ~42%, system design.

Fix the handoff before the model.

[Manu Sean @manuseen007]

@AIQuanting The harness is part of the environment, not packaging. Same weights in a different loop, with different tools and prompts, is a different agent. Small-scale version: run evals through the exact harness you ship, or the score measures an agent you don't deploy.

[Manu Sean @manuseen007]

@Didotxyz_ Delegating less while scoring higher matches what breaks in practice: every handoff is a lossy summary. Split work only where the subtask has its own acceptance check; otherwise keep it in one context and use the extra agents to review, not relay.

[Manu Sean @manuseen007]

When an agent gets a task right, I still read the trace.

Right answer, wrong path is the failure that bites next week: it skipped a check, guessed a field, got lucky.

My rule now: review how it got there, not just what it returned. Lucky passes don't count.

[Manu Sean @manuseen007]

@muddassirism The promotion holds only if someone owns the exceptions. Agents clear the routine invoices; the human job becomes defining what counts as an exception and signing off on those. If that rule isn't written down, the agent decides it by default.

[Manu Sean @manuseen007]

@RykerPhoenixLLC The agent before it writes it, against a fixed template: goal, done-when, constraints, links to sources. A person reviews the template, not every brief. If the next agent has to ask for the transcript, the brief failed.

[Manu Sean @manuseen007]

Same question: ~300 tokens of what matters vs ~113k of full chat history.

In Chroma's context rot tests, every model scored higher on the short one.

Popular advice: give the agent everything.

It can hold it. It can't use it.

My rule: agents get a brief, not the transcript.

[Manu Sean @manuseen007]

@HowToPrompt__ This is the pattern that works: the agent proposes, a verifier that can't be talked into anything decides. Same rule for everyday code work. If you can't write the check, you don't have a task for an agent yet.

[Manu Sean @manuseen007]

A lot of the "agent" jobs I sketch end up as a plain script.

If the steps are the same every run, a model only adds cost, latency and a new way to fail.

My test before adding an agent: does this step need judgment?

No: write the script.
Yes: agent for that step, code around it.

[Manu Sean @manuseen007]

@HochstatMichael The fix I keep landing on: write the acceptance checks before the agent starts, and have the checker run them against sources, not re-read the draft. A second agent grading prose is just more prose. Pass = every requirement verified, not "looks done."

[Manu Sean @manuseen007]

@nishtrivedi Agreed. The "reports to" line is the one people skip. When an agent misfires at 2am, the registry should tell you who gets paged and who can flip the off switch, not just what broke.

[Manu Sean @manuseen007]

1 in 19 tech leaders can't say how many AI agents are running in their org (Komodor survey, n=1,000).

27.9% run more than 100 ops agents.

Building my agent team, I keep one registry: each agent, its job, what it can touch, its off switch.

Not on the list? It doesn't run.

[Manu Sean @manuseen007]

@CodingDevTools Seen the same thing. Two memories, both "true", agent picks whichever ranks higher. What fixed it for me: every decision record gets a status (active or superseded) plus a pointer to what replaced it. Retrieval skips superseded by default, history stays searchable.

[Manu Sean @manuseen007]

@Yash_KVR Good list. One layer under it: don't give the agent creds that can do these at all. A denylist always misses command #8. Ask-before-run for anything irreversible, and scoped tokens so prod simply isn't reachable from the agent's shell.

[Manu Sean @manuseen007]

@rmgn Exactly, the reflex is patch, not dissolve. A line in the prompt fades after a few turns. What sticks better is making it a required output: list what could be removed before any fix is allowed. Then "delete it" becomes a valid answer.

[Manu Sean @manuseen007]

@Stellar_Civi Yes, but as a weekly budget cap, not just a number someone reads later. When an agent hits it, it pauses and asks. A cost column nobody acts on is just a receipt.

[Manu Sean @manuseen007]

@nitishmutha Agree, and the piece that breaks first is ownership. When I give an agent a role instead of a task, it needs three things written down: what it owns, who it hands off to, and what it can't do without sign-off. Tasks fit in a prompt. Responsibilities need an org chart.

[Manu Sean @manuseen007]

Agents retry. Timeouts, rate limits, a flaky API.

So any step that isn't safe to run twice will eventually run twice: the double email, the duplicate invoice.

Rule I'm building in: every agent action checks before it writes, or carries an idempotency key.

Assume the retry.

[Manu Sean @manuseen007]

@superalesha The irony is less interesting than the line they drew: by job title, not by permission. A coding agent gets the same shell and network access whoever launches it. Scope what the agent can touch and log it, and who runs it matters a lot less.

[Manu Sean @manuseen007]

@SepasZohari Balanced, not proportional. Each action type gets a minimum number of reviewed runs per week, so rare ones get oversampled. For rare and high-risk ones (sends, spends, deletes) I don't sample at all: every run gets a human until it builds a track record.

[Manu Sean @manuseen007]

52% of enterprises that test AI before release shipped an agent that passed, then failed customers (VB Pulse, Sep 2026).

"Write better evals" isn't the fix. A test set only covers what you thought of.

My rule now: first real runs go past a human, and every miss becomes a test.