Know What Your Agent Is Good At

The Build Loop, same as every rung — spec, direct, checkpoint, verify. Before we sharpen any single turn, one thing shapes all four: knowing who you're directing.

You are already an expert at this. You've just never had to do it for a worker this strange.

Every time you hand off a piece of work, you make a silent judgment about who's receiving it. You know which colleague to trust with the fiddly data reconciliation and which one to trust with the client-facing narrative. You know who needs the brief spelled out and who fills the gaps well. You do this so automatically you've stopped noticing it's a skill. It is the core skill of getting work done through other people, and it transfers directly to getting work done through agents — with one adjustment, because the agent has a shape unlike any colleague you've managed.

The agent's shape

An agent is, at the same time, the most capable and the most limited coworker you will ever direct. Both halves are extreme, and the whole discipline is knowing which half you're leaning on.

What it's genuinely great at:

• It never tires and never gets bored. The tenth revision gets the same care as the first. The tedious pass over four hundred rows — the work a person quietly resents — is work the agent does cheerfully, at speed, every time.
• It follows patterns superbly. Show it one thing done right and it will produce fifty more just like it (this is the whole of the previous rung).
• It's a fast, fearless first-drafter. A blank page costs it nothing. It will give you something to react to in seconds, and reacting is easier than creating.
• It holds a lot in view at once. Point it at a pile of material and ask what's inconsistent, and it can hold all of it at the same time in a way you can't.

What it's genuinely bad at — and this is the important list:

• It has no memory. When the session ends, it forgets everything, completely. Tomorrow's agent is a stranger who has never met you. (Rung ten is entirely about this.)
• It has no taste. It cannot tell your best work from your worst; it treats whatever's in front of it as the standard.
• It would rather guess than admit it doesn't know. Ask for something it can't do, and it will often produce a confident, plausible, wrong answer instead of saying "I can't." This is the failure that catches people off guard, because a human colleague would just tell you.
• It's relentlessly literal. It builds what you said, not what you meant. It will not infer the obvious caveat you left unspoken, because to the agent nothing is obvious.

You don't fix the shape. You staff around it.

Here's the part that turns this from an intimidating list into a working method: you are not trying to fix any of these weaknesses. You couldn't if you wanted to. You do exactly what a good manager does with any team — you match the work to the worker.

You hand the agent the parts it's built for: the volume, the repetition, the tireless first draft, the pattern-matching, the never-complaining pass over the boring data. And for the parts it's bad at, you don't avoid them — you scaffold them. The agent has no memory, so you keep the memory in files. It has no taste, so you supply the standard (rung five). It would rather guess than admit ignorance, so you build in the checkpoint where you catch the guess before it costs you (rung seven). It's literal, so you learn to say the caveat out loud.

Every later rung in this series is, underneath, one of these scaffolds. This rung is the one where you learn to see the shape they're built around — because once you can see it, the rest stops feeling like a list of rules and starts feeling like managing a specific, knowable coworker.

That's the reframe worth keeping: directing an agent well is not a technical skill you lack. It's a management skill you already have, applied to the most literal-minded, tireless, amnesiac, taste-free brilliant coworker you will ever staff.

Staff the work like you'd staff a project.

---

→ The companion — In Tropo: Know What Your Agent Is Good At — has you hand one small job to an agent two different ways, watch its shape show up in the results, and write your own short "what my agent is good at" notes you'll reuse on every build.

Power Play

Want to take this further? Make the agent scout its own weak points before it starts. When you hand off anything that matters, add this to the end of your instruction:

"Before you begin, tell me the three places you're most likely to get this wrong, make an assumption I haven't confirmed, or produce something that looks right but isn't. Don't start building until I've responded."

A capable agent will hand you a genuinely useful list — the ambiguous requirement, the guess it was about to make, the edge case you forgot. You've now converted its worst habit (guessing silently) into its most useful one (flagging the guess out loud), and you fix the problems while they're still free to fix. Do this often enough and you learn your particular agent's shape faster than any guide could teach it.

Terms worth knowing

• Session — one continuous working conversation with an agent. It has a beginning and an end; when it ends, the agent's working memory is gone unless you saved it to a file.
• Context — everything the agent can currently "see": your instructions, the files in front of it, the conversation so far. Its judgment is only as good as what's in its context.
• Scaffolding — a support you put around a task so an agent's weakness doesn't sink it: a supplied example, a saved note, a checkpoint. You'll build one on nearly every rung.