A year ago, “assign a ticket to an AI and get a pull request back” was a demo that made people lean forward in their chairs.
Today it’s a checkbox.
GitHub lets you assign an issue to Copilot, Claude or Codex. Linear starts coding sessions on Claude Code and Codex straight from an issue. Jira lists Rovo and third-party agents as assignees, right next to your colleagues. OpenAI open-sourced Symphony, which turns every Linear issue into its own Codex workspace.
If your tracker doesn’t do it yet, it will by spring. The core loop I spent a long time building is now a free feature in tools your team already pays for.
I’m fine with that. It clarifies what actually matters.
The loop was never the hard part
Watch what happens when a team switches the loop on.
The first week is exciting. Small bugs disappear. Somebody posts a screenshot of a PR that was opened while they were at lunch.
The second week, the questions start.
Who wrote that ticket? It said “fix the export” and the agent fixed a different export. Why did it touch the billing module? Who approved the dependency upgrade it slipped in? It says the tests pass - but the tests it wrote only check that the function returns something. Who is going to review fourteen PRs before Friday? And what did all of this cost?
None of these are capability problems. The models are good. They write decent code. The problem is that nobody owns the work around the code.
Five things nobody sells you
When I look at teams that have agent licenses but no AI teammate, the gap is always the same five things.
Intake. Someone has to turn “customers complain the export is slow” into small stories with acceptance criteria and a concrete way to show each one works. That’s a product owner’s job. An agent that receives a vague issue produces a vague PR.
Policy. What may the agent touch? Which repos, which paths, how much may it spend per story, may it merge on its own? Today the answer lives in someone’s head, or in a prompt that nobody reviews.
Verification. “Tests pass” is not “it works”. If the agent writes the code and the tests, it grades its own homework. Someone has to check the result against the requirement, not against the agent’s opinion of the requirement.
Accountability. Who did what, on whose approval, at what cost? When the auditor or your CTO asks, “the AI did it” is not an answer.
Process fit. Your team has a tracker, a CI, branch protection and review rules. An AI teammate that needs its own board and its own workflow is one more tool to babysit.
Model vendors sell capability. Tracker vendors sell their own agent inside their own tool. Nobody sells the layer in between: the one that runs an AI teammate under your team’s rules, in your tools, and shows its work.
What kanman does about it
That layer is what kanman is now.
You paste a requirement or tag kanman in Slack. kanman writes the stories, with acceptance criteria and a way to demonstrate each one, and you approve them before anything lands in your Jira or GitHub.
Your policy decides what kanman may touch, spend and merge. When a call is bigger than its authority, it comes to you with a recommendation and a few concrete options. You answer in the inbox or in Slack.
Claude Code or Codex does the coding in a sandbox. kanman picks the model by complexity, so a typo fix doesn’t burn the budget of a refactor.
Before anything reaches review, an acceptance spec runs against a fresh environment. The spec was written before the code, and the agent can’t edit it. No spec, no Ready. No green run, no review. I wrote more about that in the next post.
Then you get a pull request with the evidence attached. Your people review and merge, or your policy does. Every step lands in an audit log you can export.
Your tracker stays. Your CI stays. Your review rules stay.
Why this isn’t a race against GitHub
People ask whether GitHub or Atlassian will just build this. Parts of it, sure. But each of them builds it for their own tool and their own agent. A team on Jira and GitLab, with a Claude contract and two developers who prefer Codex, doesn’t want three vendors’ opinions of what an AI teammate is.
And a lot of teams I talk to can’t send their code to whichever cloud their tracker vendor picked. That’s why kanman runs in the EU, with coding sandboxes in Frankfurt, and why the runner can sit inside your own network.
The boring part is the product
Nobody gets excited about intake templates, policy tables and evidence packs. They are boring the way seatbelts are boring.
But they decide whether an AI teammate is something you trust with real work, or something you demo once and quietly switch off.
The loop is a commodity. The operating model isn’t.
Want an AI teammate that works inside your tracker, under your rules, and proves every change? kanman - Pilot from €5,000 for 6 weeks, Team €990 per AI team per month.
Marco Kerwitz
Founder of kanman.ai