Kanban boards

A kanban board holds tasks you write yourself and agents that take them. Each board has five columns:

Column What it means
Not ready Tasks you're still writing. No agent takes them.
To do Ready. The work agent takes them in the order they're in, when it can.
In progress The work agent is doing them, or they came back from review.
Review The review agent checks the work: it approves it or sends it back with changes.
Done Approved, with a summary of what was done and how well.

The cards move on their own as the agents work. Each task keeps a log of every step and who took it.

Kanban boards are part of AI Mode: turn it on to see them.

Create a board

Click Kanban boards in the activity bar, press Ctrl+Shift+U (⌘⇧U on macOS) or open the ≡ menu → Kanban boards, then + (New board). In General:

  • Name and Icon.
  • Folder: where the agents work, usually your project's repository.
  • How tasks run:
    • Sequential: one task at a time, in the folder itself.
    • Worktrees: each task works on its own branch, in its own folder (a git worktree), so several run at once. Tasks at once sets how many, and Base branch where their branches start from and merge into. It needs the folder to be a git repository.
  • Review rounds: how many times review may send a task back before it waits for you (3 by default).
  • When the agents ask you: Never: they decide alone, Only what's essential or When in doubt. Their questions show like any agent's.

Then set its agents in Agents (below) and click Create board. The board opens in a tab; the panel lists every board with whether it's on, how many tasks are done, and a dot while an agent works on one.

Set up its agents

Each active column has its own agent. Open Agents in the board's tab (or Board settings → Agents). Along the top, one card per agent shows its column, engine, model and what it may do. Click a card to edit that agent below; each agent keeps its own color.

  • Work agent: takes the tasks in To do and does them.
  • Review agent: checks the work against the task's specification and success criteria. It approves it, or asks for changes and the task goes back to In progress with its notes.
  • Summary agent (Optional, Add a summary agent): writes the final summary of each finished task. Without it, the summary comes from the review.

For each one choose:

  • Engine and model: the Engine (Claude Code, Codex, Gemini CLI, Copilot CLI or Grok) and its Model.
  • Permissions: What it may do (only look, edit files, or also run commands) and whether it Can search the web. Work and review agents start able to run commands, since doing the work and checking it both take installing, building and running the tests. With less, the agent gets stuck as soon as it needs a command.
  • Instructions: the Context, what the agent should know about the project, its conventions, where things are; and the Harness, what it does with each task in this column, step by step. Each column starts with a sensible harness you can change.

The column agents are saved as regular agents, so their runs and costs show in their history like any agent's.

Write tasks

Click + New task in the board's tab, or + on the Not ready or To do column. From anywhere, Alt+Shift+N (⌘⌥N on macOS) opens a new task on the board in the focused tab, or else on the one you last opened; with several boards and none of them open, it shows the panel so you can pick one. The command palette also has New task on … for each board. A task has:

  • Title.
  • Specification: what it's about, with everything an agent needs to do it alone.
  • Success criteria: how to tell it's done.
  • Tags: short labels like bug or frontend. Type one and press Enter (or a comma); the board's other tags are offered below. Each tag keeps its own color on every card.
  • Depends on: tasks it waits for. It isn't taken until they're all done, and its card shows Waiting on #2.
  • Starts in: Not ready, to keep working on it, or To do, to let the agent take it.

Click a card to edit it at any time. An agent working on it reads the changes in its next run.

Switch it on

The On / Off switch in the board's tab decides whether its agents take tasks. Off, nothing new starts; what's running finishes.

The agents report their work and their verdict to Termo through its MCP server. If it can't start, usually because another open Termo holds its port, the board says so and takes no tasks until it can. Close the other Termo and the board carries on by itself. With Keep agents running when Termo closes (see Local agents), a board that's on keeps working with Termo closed.

Follow it

Each card shows its title, the start of its specification and its tags. Its footer says whether it has success criteria, how many tasks it depends on, how many log entries it has, what it has cost so far, and when it last changed.

  • A card an agent is working on glows, with a chip that says which agent and model (Work · …, Review · …) and a line with what it's doing now.
  • Next #1, Next #2… mark the order the To do tasks will be taken in.
  • Review 2/3 counts how many times review sent it back.
  • A yellow strip with Decide means it's waiting for you, and says why: Couldn't finish, No agreement after 3 reviews, Review without a verdict, You stopped it… See Decide on a stuck task. Waiting for your answer means an agent asked you something.
  • In Done, Complete, Partial or Failed says how it went.

Open a card to see its Log: every step, who took it (You, Console, Work agent, Review agent…) and when. Add your own notes there; the agents read the latest entries. Summary shows the outcome, what the work agent reported, the review's notes and the cost.

Termo notifies you when a task needs you or reaches Done.

Decide on a stuck task

No agent takes a task that's waiting for you. Click Decide on its card: the top of the task asks what to do, with the reason underneath and See what the agent did to open its last run. The choices depend on what happened, the likeliest first:

  • Try again, Review it again, Another round of work or Carry on: the column's agent takes it again where it is.
  • Back to work: from review, it goes back to In progress for the work agent to fix it.
  • Approve as it is: it moves to Done with what's been done.
  • Back to Not ready: no agent takes it until you rewrite it and move it to To do.

If the task failed and the agent that would take it again can't run commands, the top says so: that's most likely why, since installing, building or committing all take one. Let it run commands and try again changes what that agent may do on the board and takes the task up again.

Write in Instructions for the agent what it should do differently, for example leave the integration tests for later. The next agent that takes it reads them before anything else. Whatever you choose, its review rounds count from zero again.

Move cards by hand

Drag a card to another column, or up and down to change the order. Moving a task back to Not ready or To do resets its review rounds. If an agent is working on the task, Termo asks first: Cancel and move stops the agent and moves it.

To stop an agent without moving the task, open the card and click Stop agent.

Merge a task's branch

In Worktrees mode, a finished task's card shows its branch. Open it and click Merge branch: Termo merges it into the board's base branch, in the board's folder, and removes the task's worktree. The folder has to be on the base branch. If the merge conflicts, it's undone and the folder is left as it was.

Add tasks from a console

The Claude Code, Codex, Gemini CLI and Copilot CLI consoles you open in Termo can see your boards and add tasks to them. Turn on Preferences → AI Mode → Agents → From consoles → Add tasks to your kanban boards, then open the console. A console that was already open does not get the option.

Then ask it in your own words, for example "add a task to the Web app board to rate-limit the login endpoint". The console writes the task so an agent can do it alone. It always goes to Not ready: no agent takes it until you look it over and move it to To do. That way, nothing a console was led to write by what it read runs with the board's permissions without you seeing it first. A console can also list a board's tasks and edit the ones in Not ready, but it never moves them along or switches a board on. Its changes show in the task's log as Console.