Dev.to AI πŸ€– Ai πŸ‘ 0 πŸ“– 4 min read

How to hand a computer job to an AI agent (and check what comes back)

A brief that gets done right the first time, and the checks that prove it was. We build King AI, an AI agent that can do anything a human can do from a computer. That means we read a lot of task briefs, and the differen

A brief that gets done right the first time, and the checks that prove it was.

We build King AI, an AI agent that can do anything a human can do from a computer. That means we read a lot of task briefs, and the difference between a job that comes back right and one that comes back sideways is usually decided in the first three lines someone types.

Handing work to an agent is closer to delegating to a contractor than to calling a function. Here is what changes, how to write the brief, and what to check before you trust the result. None of it depends on which agent you use.

What changes when an agent does the job

You specify outcomes, not steps. With a script you write every step. With an agent you describe the end state and let it find the path. If you catch yourself writing step five of twelve, stop and describe what "finished" looks like instead.

Access is the real permission model. Whatever accounts, folders, repos and cards you share defines what can happen. Give the access the job needs, and nothing that only an unrelated job would need. A read-only export beats a live admin login when reading is all the task requires.

"Done" has to be checkable from outside. A human contractor's "done" comes with context you can ask about. An agent's "done" should come with evidence you can open: a URL, a file path, a message id, a commit.

Some actions only go one way. Sending, paying, publishing and deleting cannot be taken back with an undo. Decide in the brief which of those the agent may do on its own and which need your yes.

Writing the brief

A brief that works has six parts. Most of them are one line.

Outcome:      What exists or is true when this is finished.
Done means:   How you (or anyone) can check it, concretely.
Inputs:       Where the files, data, logins and examples are.
Constraints:  Budget, deadline, tone, format, tools to use or avoid.
Ask me before: The one-way actions that need a yes.
Report back:  What evidence to hand over (links, ids, paths).

A vague brief:

Clean up our customer list.

The same job, briefed:

Outcome: one deduplicated customer sheet from customers_2026.csv.
Done means: one row per email address; phone numbers in +1 format; a second tab listing every row you merged or dropped and why.
Constraints: never delete a row outright, move it to the second tab. Dates in the source are US format (03/04 means March 4).
Ask me before: nothing, this is read-and-write on a copy.
Report back: the file path and the before and after row counts.

The second version takes a minute longer to write and saves the round trip that the first one guarantees.

Three details cause most of the failures we see:

  1. Ambiguous formats. Dates, time zones, currencies and units. "03/04" is March 4 in one country and 3 April in another. Say which.
  2. A missing quality bar. "Make it better" has no finish line. "Load in under two seconds on a phone" does.
  3. Assumed constraints. If the site must stay on its current host, or the email must come from a specific address, write it down. The agent cannot read the assumption in your head.

What to verify in the result

Read the artifact, not the summary. Open the URL. Run the script. Open the spreadsheet. A report that says "deployed" is a claim; the page loading at its address is the fact.

Check numbers against the source. Row counts in and out, totals against the printed totals, the number of records against what you expected. If the agent merged 40 duplicates, the before and after counts should differ by 40.

Spot-check at random. Look at three rows from the middle, not the first three. Errors hide where nobody looks.

Confirm one-way actions happened exactly once. For a send, look in the sent folder or the recipient's thread. Retries are the classic way to email someone twice, so check before you ask for a resend.

For code, run it yourself. Run the tests, read the diff, and try the one input you care most about.

Ask for evidence up front. If the brief says "report back with links and ids", checking takes a minute instead of an investigation.

The brief is the interface

An agent that can use a computer the way a person does turns a lot of jobs into a sentence or two. The quality of those sentences decides the quality of the work. Write the outcome, say how to check it, name the one-way doors, and then check the thing itself.

If you want an agent built around exactly that loop, a sentence in and finished work out, King AI is opening by invitation at usekingai.com.

πŸ“° Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.