SI Agent News
DevelopingClaudeOpenAIAgents

A student argues AI agents are now something you hand work to, not build

A developer-in-training says daily work now means briefing coding agents and checking their output, and he lists four habits that make that safer.

The short version

Why it matters

If agents increasingly do the typing, the valuable skill is writing a clear brief and checking what comes back. That skill works with any vendor's tool and is useful well beyond programming. It is one student's view rather than a study, but the habits are practical and easy to test.

A shift seen from the desk

As dev.to reported in a post by a second-year student of intelligent informatics at ZEN University, the talk about AI agents has changed direction. An agent here means software that takes several steps toward a goal on its own. Not long ago, the discussion was about how to build one: which model to use, how to connect tools, how to give it memory or let it look up documents.

The author says that in his own daily work the question is now how to hand tasks to an agent. He uses Claude Code and Codex, two coding agents, every day for implementation, tests, research and documentation. His role, as he describes it, has moved from the person who writes the code toward the person who decides what to hand over and then checks the result.

He is clear that this does not replace writing code. Writing, delegating and verifying now sit side by side. He also says he will not make up figures about how much faster any job became, so the post offers habits, not benchmarks.

Brief first: the goal and the context

The first habit is to split the work and write down what finished looks like. A vague request such as "handle it well" tends to make an agent widen the scope or start improving things nobody asked about. The author suggests one sentence for the goal, a short list of done conditions, and a list of things not to do. His example is adding email and password checks to a login screen, with existing tests passing, new tests added and the rest of the code left alone.

Large requests should be cut into steps. Instead of asking to make a feature better, ask for tests first, then the fix, then the documentation update. That way you and the agent judge "is it done?" against the same bar.

The second habit is passing context. An agent may be able to read a project's files, but it does not know the assumptions that live only in your head. The author suggests telling it:

Naming exact paths that must not change, and the folders that may, also makes the final set of edits easier to review.

Check the work, then set the limits

The third habit is verification. An agent can report that tests passed while the tests cover very little. The author does not treat a result as ready to merge, meaning ready to join the main project, until he has reviewed it. He looks for new tests, for failures that were quietly hidden, and for changes outside the requested area. He also checks for unrequested rewrites, new dependencies and secrets that slipped into the change.

He will ask an agent to run the tests and report back, but he still reads the diff, the list of what changed, and runs the code locally when needed. In his words, verification is part of delegating, not its opposite. If the done conditions are not met, the missing pieces go into the next request.

The fourth habit is permissions. Making everything manual makes delegation pointless, while making everything automatic risks actions you cannot undo. His rough line: reading code, running tests, drafting a proposed change and editing on a separate working branch are often fine to automate. Forced overwrites in git (the tool that tracks a project's history), deleting files or branches, touching secrets or production settings, deploying, publishing and adding broad-permission dependencies should wait for a person.

His key distinction is between what a tool can do and what it may do. Anything irreversible, anything that leaves the machine and anything that publishes stays on the human side.

Why building still helps

The author does not say building agents has become irrelevant. He built Yata, a personal assistant that runs on his own PC. You chat with it in Japanese, and it helps with research and with checking mail and calendars. For any action that leaves the machine, such as sending mail or creating a calendar event, it waits for a person to confirm.

That experience, he argues, makes him a more careful delegator. He has seen models ignore instructions, tools fail, context run short and broad permissions trigger unintended actions. Knowing this makes it easier to decide where an agent should stop, what to tell it first and what counts as done.

His conclusion is that building and delegating feed each other. If you can build, you can design how to delegate. If you delegate, you see what is worth building next.

How much weight to give it

This is a single student's account from his own projects, not a survey or a measurement. He offers no data on time saved or errors avoided, and says so. Treat it as a well-organised description of one person's routine.

Still, the four habits are ordinary management practice applied to software: set the goal, share context, inspect the output, limit what can go wrong. Because they are plain, they are easy to try on one small, reversible task and judge for yourself.

What to watch

Try it

Pick one small, reversible task for an AI assistant. Before sending it, write the goal in one sentence, three done conditions and one thing it must not touch, then review the result yourself.

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by dev.to · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news