Dev.to AI ๐Ÿค– Ai ๐Ÿ‘ 0 ๐Ÿ“– 7 min read

Your Coding Agent Shouldn't Be Able to Close Its Own Ticket

Would you let a developer approve their own pull request? GitHub won't. Open a review on your own pull request and approving it simply isn't an option. And nobody takes that as an insult to developers. It's just how rev

Your Coding Agent Shouldn't Be Able to Close Its Own Ticket

Would you let a developer approve their own pull request?

GitHub won't. Open a review on your own pull request and approving it simply isn't an option. And nobody takes that as an insult to developers. It's just how review works: the person who wrote the change isn't the person who says it's good.

So why do so many of us let coding agents close their own tickets? ๐Ÿค”

You know how it goes. You hand the agent a ticket. It works away for a while. Then:

Implemented and verified. โœ…

And the ticket moves to Done. Sometimes because the agent moved it. Sometimes because you did, on its say-so, at six on a Friday evening. ๐Ÿ˜… Either way, the only evidence the work is finished is the word of the thing that did the work.

I think that's backwards. In this post I'll argue that an agent should be able to claim a ticket is done, with evidence, but never decide it. Then I'll show what that split looks like in practice.

Full disclosure: I make Arbituro, a work tracker built on exactly this split. The argument holds whatever tracker you use, and the Arbituro bit is near the end.

๐ŸŽฏ "Done" Is the Agent's Finish Line

Here's the thing about a coding agent: it's trying to finish. That's not a flaw, it's the job. Every step it takes heads towards the moment it can say "done" and stop.

So for an agent, "done" isn't just a status. It's the finish line. And if the agent also gets to decide where the finish line is, you get exactly the incentive you'd expect. ๐Ÿ™ƒ

None of this needs bad intent. It's what happens whenever the one being measured also does the measuring. We worked that out for people a long time ago. It's why code review exists, why auditors aren't the accountants, and why your own approval on your own pull request doesn't count.

๐ŸŽญ Confidence Isn't Evidence

The second problem: an agent reports success in the same cheerful tone whether it succeeded or not.

If you've worked with one for a while, you've probably seen some version of these:

  • the tests pass... because the agent changed the test;
  • "works on mobile", when nothing was ever opened on a phone;
  • the happy path is fixed, and the ticket was about the edge case.

Every one of them arrives as "Implemented and verified. โœ…" The sentence is identical. Only the evidence would tell them apart, and the sentence has none.

โœ‚๏ธ Split the Claim from the Decision

So here's the fix, and it's simple: two different actors, two different verbs.

The agent claims. When it believes a criterion is met, it says so, with what it checked: the command it ran, the page it opened, what it saw.

Someone else decides. A CI check, for anything a machine can verify. A named person, for anything that needs judgement.

That's the whole idea. The agent keeps all of its speed. It writes the code, opens the pull request, runs the tests and makes its case. It just doesn't mark its own work. โœ๏ธ

Here's a claim from Arbituro's demo workspace. The criterion is "Sam is happy with the wording on the account page", and the agent wrote:

Rendered the account page on the preview with a test subscription: "Add a cold brew four-pack to this delivery, $19.95" sits under Next delivery, and the confirmation reads "Added. It comes with your coffee on Tuesday."

Sam can decide that in one press. Met, or Not met with a reason. Either way it's Sam's call, and it's on the record.

Notice what the claim doesn't do: it doesn't say the wording is good. It can't. Only Sam knows whether that's what Sam wanted. ๐Ÿ’ฌ

โš–๏ธ Checks Decide What They Can, People Decide the Rest

This split doesn't mean a human clicks through every ticket. Most criteria on a code change can be decided by CI:

  • "Checkout refuses a postcode outside the delivery area" is decided by the checkout check.
  • "Every existing billing test still passes" is decided by the test check.

When the check passes on the pull request's latest commit, the criterion is met. When it fails, it isn't. Nobody presses anything. ๐Ÿค–

People are for the rest: wording, design, whether the change solves the customer's problem. Name one person per criterion. "The team" never decides anything.

Writing criteria that a check or a person can decide is its own skill, so I wrote it up separately: Writing Acceptance Criteria an AI Agent Can Actually Meet.

๐Ÿ”€ Merged Isn't Done Either

The same trap shows up one level higher. Plenty of setups close the ticket automatically when its pull request merges.

But a merge only says the code is in. It says nothing about "a four-pack can be bought on a phone", or whether the customers understood the new page, or the email that still has to go out. Those are exactly the criteria the code alone can't show.

A merge should ask the person who signs the job off to finish it, not finish it for them.

๐Ÿคจ "But Isn't This Just Slower?"

You might have a few objections by now. Here are the ones I think matter.

"It's slower." For a criterion a check decides, nobody waits on anybody: it's decided the moment the check finishes. For one a person decides, it's one press against written evidence. Compare that with the cost of a wrong "done": the bug report, the reopen, and the context everyone has lost by then.

"People will just rubber-stamp it." Some will. But a rubber stamp has a name on it now. And next to a claim with real evidence, "Implemented and verified" looks exactly as thin as it is. ๐Ÿ˜ฌ

"Agents will get good enough that this won't matter." They'll be right more often, for sure. But "done" isn't only about correctness. It's whether the work is what the person who asked for it actually wanted, and no agent gets to decide that for them, however good it gets.

"My agent works as me anyway." That's exactly the problem. Over MCP, an agent usually works with your permissions. So if you can move a ticket to Done, so can it. A line in the prompt asking it nicely not to is a request, not a permission. The fix has to live in what the tools allow. ๐Ÿ”’

๐Ÿ› ๏ธ What You Can Do Today, in Any Tracker

  1. Give every criterion a decider: a named CI check or a named person.
  2. Let CI decide whatever it can, and name the check in the criterion exactly as it shows on the pull request.
  3. Ask for evidence in every claim. Put it in your CLAUDE.md or AGENTS.md: never mark a ticket done; comment with what you checked, and who should decide.
  4. Don't let a merge close the ticket by itself. Let it ask someone.
  5. Take the button away if you can. If your agent can have an account of its own, give that account no way to move a ticket to Done.

How to wire the first three into Claude Code is in Stop Pasting Tickets into Claude Code.

๐Ÿ—๏ธ How Arbituro Does It

Arbituro is built around this split, so the agent never gets the button in the first place:

  • No closing tool. A coding agent connected over MCP can read jobs, claim criteria, propose changes, post notes, file jobs and documents, and link pull requests. There's no tool that ticks a criterion or finishes a job.
  • Every criterion names its decider: a person, a CI check, or an approved message being sent.
  • Claims go to the decider, with the agent's evidence, and they press Met or Not met.
  • A merge asks. When a pull request merges, the person who signs the job off is asked to finish it. Finishing is their verdict.
  • Changes are on the record. Changing a criterion is a proposal a person accepts. Waiving one takes a reason.

The agent is even told when to stop. Say the agent has claimed both of the four-pack job's person-decided criteria, and the checkout check has passed. This is the end of what get_job hands back, trimmed a little:

Done when:
- [Unknown] Says what's in the can in its first line (Priya Raman decides)
  Claimed on 2026-11-12 by Sam Taylor; not decided yet.
- [Unknown] A four-pack can be bought on a phone (the owner decides)
  Claimed on 2026-11-13 by Sam Taylor; not decided yet.
- [Passing] Checkout refuses a delivery postcode outside Melbourne (decided by the `checkout` check)

Next: nothing for you. Priya Raman decides "Says what's in the can in its first line"; Sam Taylor decides "A four-pack can be bought on a phone". Sam Taylor finishes BREW-26 once #48 merges. Tell the user, and stop.

"Tell the user, and stop." That might be my favourite line in the whole product. ๐Ÿ˜„ An agent that's told nothing tends to keep going.

And when a pull request does merge, this is what lands in front of the person who signs the job off:

A Your turn card for the job "Cold brew in the shop's menu and search": its pull request merged, and finishing ticks your clause, with Open the job, Open #46 on GitHub, and Finish it.

๐Ÿ Wrapping Up

The agent does the work and makes its case. A check or a person decides. That isn't distrust of agents. It's the same rule we've always applied to ourselves: whoever wrote the change doesn't approve it.

Arbituro is free for up to three people at arbituro.com.

What about you? Does your coding agent close its own tickets? Has a "done" ever turned out not to be done? And if you think I'm wrong about this, I especially want to hear why. Tell me in the comments. ๐Ÿ‘‡

๐Ÿ“ฐ Read the original article on Dev.to AI

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