How to make an AI agent prove a task is done
An AI agent's done is its opinion of the work, not proof. One line in its rules makes every done come with the file, link or number, and you open it.
An AI agent's "done" is its opinion of the work, not proof of it. Add one line to its rules so every "done" comes with the file path, link or number behind it, then open what it gives you. The quickest way to get an agent is HostNest, set up for you in about 10 minutes; the line works on it, or on an agent of your own.
Why does an AI agent say "done" when it isn't?
An AI agent stops when the work looks done to it, and nothing has checked it. Anthropic's guide to its own coding agent says that when the agent has nothing to run that tells it whether the work is done, "'looks done' is all it has to go on. In practice that makes you the test suite". It isn't one vendor's problem: the research lab Transluce found that a pre-release OpenAI model "frequently fabricates actions it took to fulfill user requests". A builder on Reddit asked an agent for a Word file and got back a save path and "Saved." They opened the path: "The file-writing tool had never been called."
Don't have an AI agent yet?
The quickest way is HostNest, a team of Claude agents. You sign up yourself, and the rest comes ready: your own server in the EU (a computer that's always on), the agent, and its wiring. It takes about 10 minutes, most of it the server being built:
- Account: your email (we send you a 6-digit code) or Google.
- Free trial: 7 days free.
- Telegram: a few taps create your bot. The bot is yours.
- Claude: while your server is built, you read four safety rules. Then you sign in to your paid Claude plan (Pro, Max, Ultra or Team) and paste back one code.
- Live: open your agent in Telegram, send it the line below, and ask it to keep it in its notes.
After that, the upkeep is handled for you: updates install themselves overnight, a failed one rolls back on its own, and the server is backed up daily. Each agent remembers what you teach it, in notes on its own server. Get your agent on HostNest
Do I need to code? No. You talk to your agents in Telegram. How long does it take? About 10 minutes, most of it the server being built. Will it do things without asking? It checks with you before anything goes out or costs money, unless you tell it it doesn't need to. Is my stuff safe? It runs on your own private server in the EU, with your own Claude login, and it's backed up daily.
Already running an agent of your own? Add the line to its rules file, as below.
What line do I add to the agent's rules?
Add this line to the rules file the agent reads at the start of every session (in Claude Code, that's CLAUDE.md):
Never report done on its own. Report it with the file path, link or number that proves it.No rules file yet? Give an agent one job in five lines of CLAUDE.md starts one. The same line ends the card in how to split work between AI agents.
What counts as proof?
Proof is something you can open or check yourself, matched to the job:
- A file or a report: its path, and its first lines.
- A page, a post or a message: its link.
- A number: where it came from, so you can check it at the source.
- An email or a booking: the sent copy or the confirmation.
- A change to code: the test it ran and what came back. Simon Willison sets the bar for that test: "That test should fail if you revert the implementation."
Do I still have to check?
Yes, and checking proof is quicker than redoing the work. Anthropic's guide gives the reason: "reading the output is faster than rerunning the check yourself. For a session you weren't watching, it's also the only evidence you have." Open the path, click the link, read the first lines. A file proves something was written, not that it's right.
Will the rule get lost in a long chat?
Not if it lives in the rules file. Claude Code's memory docs say the main rules file "survives compaction": when a long chat is compacted, Claude reads the file again from disk. A rule typed only into the chat can get lost.
Can I make the check automatic?
For code, yes, with a Stop hook. In Claude Code, a Stop hook is a small script that runs when the agent tries to finish; if your check fails, it blocks the "done" and sends the agent back. Anthropic's guide explains why it's stronger than a rule: "CLAUDE.md is for guidelines, and Claude is allowed to decide a guideline doesn't apply." One limit: by default, after eight blocks in a row, the agent is allowed to stop anyway.