How to Delegate Work to an AI Agent (Without Babysitting It)

Q: How much detail should I put in a brief to an AI agent?

Define the outcome, the constraints, and what 'done' looks like — then stop. If you find yourself scripting the steps, you're doing the agent's PLAN stage for it, and you'll cap the result at whatever you could think of yourself. Brief it like a capable hire, not like a macro.

Delegating to an AI agent works the same way delegating to a strong hire works: pick a job with a definable outcome, brief the outcome instead of the steps, give the agent access to the tools the job requires, and require a report when it's done. If you're watching it work, you haven't delegated — you've just changed screens.

The founders who get real leverage out of agents aren't better prompters. They're better delegators — a skill most $5–50M founders already have from building a team, and then mysteriously abandon the moment the worker is software. Here's the method, in six steps, using Harry — the Optimus background worker — as the working example.

Step 1: Pick jobs with a definable "done"

Not every job belongs in the queue. The test is one question: can you describe what "done" looks like in two sentences? "Migrate the archive and verify every record made it" passes. "Make our marketing better" doesn't — that's a strategy conversation, not a delegation.

Your best candidates are sitting in the backlog you've been carrying for months: the report that should exist, the cleanup that never happens, the integration you keep re-scoping. Multi-step, outcome-shaped, no live judgment calls in the middle. (There's a whole playbook for working through that pile: how to clear your "someday" backlog with an AI agent.)

Step 2: Brief the outcome, not the steps

The most common delegation failure is scripting: "first open the spreadsheet, then filter by date, then…" Do that and you've capped the result at whatever you could think of — and made yourself the author of every future failure. A background agent runs its own PLAN stage; that's its job, not yours.

A good brief has four parts:

Outcome — what exists when this is done.
Context — what the agent needs to know about your business to do it right.
Constraints — what it must not touch, spend, or send without you.
Definition of done — how you'll both know it's finished.

Step 3: Give it reach into your actual stack

An agent that can't touch your systems can only produce advice. The work lives in your CRM, your inbox, your data, your platforms — so the agent needs wired-in access to them, or every job ends with you doing the last mile by hand. Every Optimus agent connects to the tools you already run through one secure gateway, scoped to your own keys (a patented approach), so the reach is real and the keys stay yours. If access is the thing making you hesitate, read the straight answer: is it safe to give an AI agent access to your business tools?

Step 4: Demand a report-back contract

This is the step that kills babysitting. Before you trust an agent with anything real, it has to guarantee that every job ends in a message to you — either "here's what you wanted" or "here's the exact next step." No silent dead ends, ever.

Harry's whole protocol is built around that contract: PLAN → EXECUTE → TROUBLESHOOT → HEAL → REPORT. When a step breaks, he tries alternatives, repairs the underlying workflow, and then reports. The only time you hear from him mid-job is when the wall genuinely needs your judgment — and then you get a specific, answerable question, not a shrug.

Step 5: Let the failures make the system better

With a human hire, mistakes cost twice: once in the error, once in the retraining. With a self-healing agent, a failure is a one-time purchase. Harry fixes the underlying skill when something breaks, so the next run is cleaner and the same failure doesn't recur. Over months, that compounds into a workforce that gets sharper on your specific business with every job it runs. (The mechanics are in what a self-healing workflow is.)

The practical implication: don't over-protect the first runs. Give real jobs, review the reports, tighten the briefs. Early friction is the system learning your stack.

Step 6: Review outcomes, not activity

You wouldn't stand behind a senior hire watching them type, and the same discipline applies here. Judge the deliverable when the report lands. If the output misses, fix the brief — a vague "done," missing context, a constraint you forgot — and re-dispatch. Managing agents by watching them work is the software version of micromanagement, and it burns the exact hours you were trying to reclaim. It's mistake #6 of the seven delegation mistakes founders make.

What stays yours

Everything irreversible. Pricing, firing, signing, sending the email that can't be unsent — the agent can prepare all of it, but the trigger stays under your finger. That's the architecture: agents do the work, the architect decides what's worth doing. Delegate the three hours of execution; keep the thirty seconds of judgment.

One more unlock once the method is habit: you stop needing a desk to delegate. Founders running the full Optimus crew hand jobs to the system from their phone via Mako on the go, and Harry has the result waiting before they're back at a keyboard.

FAQ

How much detail should I put in a brief to an AI agent?

Define the outcome, the constraints, and what "done" looks like — then stop. If you find yourself scripting the steps, you're doing the agent's PLAN stage for it, and you'll cap the result at whatever you could think of yourself. Brief it like a capable hire, not like a macro.

What if the agent gets stuck halfway through?

A real background agent never leaves you guessing. Harry's contract is: you get back either "here's what you wanted" or "here's the exact next step." A stuck job comes back to you as a specific, answerable question — not as silence you discover weeks later.

Should I watch the agent while it works?

No — watching defeats the point. Delegation pays when the work leaves your attention entirely. Judge the output when it reports back, the way you'd judge a contractor's deliverable. If you can't resist watching, the job's definition of done wasn't sharp enough.

What work should I never delegate to an agent?

Irreversible decisions: firing, pricing, signing, anything you can't take back. Agents execute; the architect decides. Delegate the three hours of execution, keep the thirty seconds of judgment.