Rewritten on 1 October 2026. The first version of this post described a model with several AI teammates on one board. kanman has since become one AI teammate per workspace, working inside your team’s tracker. The argument below is the same; the product it describes is the current one.
“Explain this stack trace.”
“Write a test for this function.”
“Refactor this file to use the new API.”
This is the assistant paradigm. You ask, it answers. You drive, it types. Every developer on your team has one by now: Copilot in the editor, Claude Code or Codex in the terminal.
And it works. For a lot of things, it works very well.
Assistants have their place
Let’s be clear: coding assistants are genuinely useful.
A developer debugging at 2am doesn’t need a process. They need something that reads the error, suggests a fix and writes the boring utility function. One person, one tool, direct interaction. No overhead.
None of that goes away. kanman doesn’t replace your assistant. Keep using whatever your people like in their editor.
But assistants don’t own anything
The assistant paradigm hits a ceiling as soon as the work is bigger than one person’s session.
An assistant puts all the load on you. You decide what to ask. You phrase it. You check the output. You remember what’s left. When you close the terminal, the assistant forgets the work existed.
An assistant never says “this requirement is ambiguous, here are two readings, which one do you mean?”. It never notices that the story it’s working on overlaps with one from last week. It doesn’t know what it’s allowed to touch, because nobody told it, and it can’t prove it’s done, because “done” is whatever it thinks.
Today, every developer on your team has an assistant. Nobody has a teammate.
What a teammate does differently
Think about a good colleague.
They take a vague request and come back with a plan you can say yes or no to. They know the rules: which parts of the codebase are touchy, what needs a second pair of eyes, how much time a thing is worth. When something is above their pay grade, they ask, and they come with a recommendation instead of a bare question. When they say it’s done, they show you.
That’s what kanman is now: one AI teammate per workspace. It says “I”. It works next to your people, not in a separate tool.
- It takes requirements. You paste one, or tag kanman in Slack. I turn it into stories with acceptance criteria and a concrete way to demonstrate each one. You approve them before they land in your tracker.
- It follows your rules. A policy decides what I may touch, spend and merge. Bigger calls come to you with a recommendation and a few options.
- It shows its work. Before anything reaches review, an acceptance spec runs on a clean checkout. The pull request arrives with the evidence attached.
- It’s accountable. Every decision, approval and cost lands in an audit log you can export.
The coding itself is done by Claude Code or Codex, in a sandbox. Same agents your developers already know. The difference is everything around them.
Same tracker, same rules
This part matters more than it sounds.
When AI works in its own interface, your team ends up maintaining two worlds: the real backlog in Jira or GitHub, and the AI’s version of it somewhere else. Context drifts. Nobody knows which one is true.
kanman works where your work already is. Stories go into your Jira or GitHub Issues. Code goes through your CI, your branch protection, your review rules. kanman shows up in the activity feed like any colleague, with the same avatar treatment, the same comments, the same history.
No second board to keep in sync. No new workflow to learn.
Autonomy is a dial, not a switch
“But I don’t want AI deciding things without me.”
Fair. Neither do I, for most things.
That’s why autonomy in kanman is a policy, not a vibe. Presets range from Trial, for the first week on a repo without specs, to Autonomous for mature teams. Every decision is classified: some I just do, some I do and tell you, some I always ask about first. Risk can only push a decision towards “ask first”, never away from it. And budgets stop a run before it gets expensive, instead of after.
Most teams start careful and loosen up as the evidence piles up. Some decisions stay with humans forever. The point is that you choose, and you can see what was chosen.
The shift worth making
Assistants were the first generation. They proved the technology works.
A teammate is the next step. Not a smarter autocomplete, but someone who owns work end to end, inside your rules, and shows you the proof.
Your team doesn’t need more assistants. It needs one teammate it can trust.
Ready to move beyond assistants? Meet the AI teammate that takes requirements, follows your rules and proves its work. kanman - Local free, Starter €99 and Team €249 per AI team per month, 14-day trial without a card.
Marco Kerwitz
Founder of kanman.ai