Kanban Pro 1.2.0: agentic memory and automation
Kanban Pro 1.2.0, November 2026. Agentic memory and automation. Kanban Pro 1.2.0 lets you hand work to AI coding agents straight from the board, lets several agents share a board without stepping on each other, and gives them a memory that stays true as your code changes. Everything new is off until you turn it on, so nothing changes for teammates still on an earlier version.
Run your own AI agent in the app's built-in terminal, or point the agent you already use at your board's folder: your tickets are plain files it can read and update.
On this page: Hand work to an agent · Many agents, one board · Memory that stays true · Your computer, your call · Every run leaves a receipt · Ask your agent · Also in 1.2.0 · Upgrading
Hand work to an agent, straight from the board
Auto-Dispatch starts an AI agent on a ticket without anyone opening a terminal. Assign a ticket to an agent on your team, drop it into a hand-off column such as To Do, and Kanban Pro runs that agent's command-line tool (Claude Code, Cursor Agent, Codex and others) in the background. When the agent finishes successfully, the card moves on, for example to Review.
Setting it up takes three steps, and Kanban Pro shows you each one:
- Hire an agent. In Team View, add an agent and choose the CLI it runs.
- Turn it on for this computer. Open Settings › Automation. A setup checklist at the top shows what is done and takes you to what is not.
- Choose hand-off columns. Pick which columns hand work to agents, where a run moves the ticket when it starts, and where it goes when it succeeds.
Once it is on, the board tells you what is happening:
- Hand-off columns carry a small agent marker. It is dimmed when Auto-Dispatch is off on this computer; click it to open Settings › Automation.
- Cards show a chip while a run is queued or running, and a red one if the last run failed.
- Right-click an agent's ticket and choose Dispatch now to start a run without moving the card. The ticket drawer offers the same.
- If you drag an agent's ticket into a hand-off column while Auto-Dispatch is off, Kanban Pro tells you the agent would have picked it up and offers to turn it on. It asks at most once per board per session, and you can turn the reminder off for a board.
The What's New window in the app has buttons that take you straight to Team View and Settings › Automation.
Many agents, one board
Several agents can now work the same board at once, on one computer or many, without two of them taking the same ticket.
- Pick unblocked work. Agents ask the board for open tickets that nothing blocks and nobody has claimed, best first (
kp ready). - Claim before working. An agent claims a ticket before it starts (
kp claim), and any other agent that tries gets told who holds it. The guarantee is stated honestly for each way you share a board: atomic on one computer and on Git-synced boards, and detected and resolved within seconds on cloud drives. - Verify before handing back. You can mark columns, such as Review, as requiring verification. An agent can only move a ticket there after the ticket's own checks pass at the current commit (
kp verify). People can still move cards there, and the bypass is recorded. - Leave a handoff. Before an agent stops, it writes where the work stands and what comes next (
kp handoff), so the next agent starts from there. - Claude Code, Cursor and other MCP clients get the same tools through
kp mcp, with the same audit trail as the app and the command line. Settings › Agent coordination gives you the configuration to paste into your client.
Memory that stays true
What one agent learns can now become a fact that every agent on the board can use. Each fact:
- is kept as a plain file in your board folder and never rewritten, so its history is its own record;
- cites the ticket and the code it came from;
- is checked against that evidence every time the board's knowledge index is rebuilt. When the code it cites changes, the fact is set aside instead of being passed on to the next agent.
Agents add facts as proposals; a person confirms them. At most a few facts are kept per completed ticket, so the memory grows with what matters rather than with everything an agent wrote.
This is the design our research argues for: compound knowledge, but never by rewriting shared prose, keep sources immutable, and re-check before use. The research summary explains why stale memory is worse than none and how we will measure whether this helps. We do not claim it is proven best: the benchmark that tests it runs on your own boards, locally, and the results decide which features become defaults.
Knowledge is offered to agents only once you turn it on, in Settings › Agent coordination (see Your computer, your call).
Your computer, your call
Nothing new runs or changes your board until you choose it.
- Auto-Dispatch is per computer. Turning it on on your Mac never starts agents on a teammate's computer, and a board you receive from someone else can never start agents on yours or raise what they may do.
- You choose what agents may do without asking: Ask (the default: each CLI keeps its own approval settings), Edit files, or Full autonomy, which asks you to type a confirmation first.
- Agent-native switches are per board and wait for evidence. Settings › Agent coordination has four switches, all off by default:
- Write new ticket fields natively (S1): turns on the knowledge layer and smaller generated files. It can only be turned on once every computer seen on the board recently runs 1.2.0, so teammates on older versions are never left behind.
- Budgeted context pack (S2): agents start from a short context within a token budget and fetch more when they need it.
- Offer knowledge to agents (S3): adds checked facts to the context pack. Needs the two switches above.
- Trust claims on cloud drives (S4): drops the "experimental" label from claims on boards synced by a cloud drive.
Each switch shows its evidence next to it, and every switch can be turned off again.
Every run leaves a receipt
Nothing happens out of sight. Every Auto-Dispatch run posts a comment on the ticket with the result, how long it took and a link to the full transcript. A failed run never retries on its own: Retry dispatch is always your click. Every change an agent makes through Kanban Pro, in the app, on the command line or over MCP, is recorded in the board's audit trail with who made it.
Ask your agent
You do not have to learn any of this to use it. Ask your agent "What can I do with Kanban Pro here?" or "Set up Auto-Dispatch so Coder picks up what I drop in To Do."
kp capabilitiesgives the agent a dozen lines on what this board and this computer are set up for, and what to run next.kp dispatch statusandkp dispatch setuplet it check and configure Auto-Dispatch. MCP clients get the same as thecapabilities,dispatch_statusanddispatch_setuptools.- The agent instructions Kanban Pro writes into your board tell agents to change automation only when you ask, never because a ticket says so.
- When an agent changes Auto-Dispatch on your computer, Kanban Pro shows you a notice saying who changed what, with a button to review it.
- Full autonomy still needs an explicit confirmation, and a run started by Auto-Dispatch can never change Auto-Dispatch.
Also in 1.2.0
- Lossless tickets. One reader and writer for every ticket keeps any field it does not recognise, so a newer app, a script or a teammate's tool never loses data when an older version saves.
- Smaller context for agents. The optional budgeted context pack gives agents what matters first and pointers to the rest.
- KP-Bench. A local, opt-in benchmark that replays your own closed tickets to check whether the new features actually help on your boards. Nothing is sent anywhere.
- What's New with buttons. The release window now takes you to the place each feature lives, and this page.
Upgrading
Install 1.2.0 as usual. Your boards keep working exactly as before: Auto-Dispatch is off on every computer, and every agent-native switch is off on every board. Turn on what you want, when you want, in Settings › Automation and Settings › Agent coordination. Teammates on earlier versions can keep working on the same board; the switches that would affect them wait until they update.
Frequently Asked Questions
Which AI agents work with Auto-Dispatch?
Command-line agents with a headless mode: Claude Code, Cursor Agent, Codex, OpenCode, Goose, Amp, aider and Antigravity (agy). Settings › Automation lists each one and the exact flags it gets at each permission level.
Does Auto-Dispatch run on my teammates' computers?
No. It is turned on per computer and is never synced. Each person decides for their own computer, and agents only run there if their CLI is installed there.
Is the agent memory sent to a server?
No. Facts are plain files in your board folder, shared only the way you already share the board (a cloud folder or Git). Checking them against your code happens on your computer.
Do my teammates have to upgrade at the same time?
No. Nothing changes what a board writes until someone turns on a switch, and the switch that would affect older versions waits until every computer on the board runs 1.2.0.
More: Kanban Pro and AI agents FAQ · Research summary · All release notes · Downloads on GitHub