What columns should a kanban board for AI agents have?

At minimum: a not ready state, an active state, a blocked state, a review state for unreviewed agent output, a done state and a place for decisions waiting on a person. Liquid Brain uses nine states, from backlog through declined [C2].

Does every card have to pass through every state in Liquid Brain?

No. The states are available options, not a mandatory linear sequence [C2]. A card can sit wherever its real status is. VERIFY marks work that still needs review, and a decision waiting for a person is different from active implementation, so each state tells you who has to act next.

Can I limit how much work runs in parallel?

Yes. Liquid Brain lets you set a project WIP limit and opt into explicitly weighted completion using a chosen baseline [N08]. Only DONE earns completion weight, and missing weights or changed scope can make weighted progress unavailable, so treat the figure with care and read it alongside the board.

kanban for AI agents: how an AI project harness handles it

Kanban for AI agents: the board states that matter when software does the work, including review, blocked and decision owed.

Kanban for AI agents is ordinary kanban with one adjustment: the worker is software, and software is bad at telling you when it should not have started. A human colleague who picks up a blocked ticket will usually notice and ask. An agent given the same ticket will attempt it. So the board stops being a communication aid and becomes a control surface. Which column a card sits in decides whether anything is allowed to act on it, and the columns therefore need to be more precise than the classic to do, doing, done. Teams that adopt an agent often discover they need to distinguish work that is waiting for a human decision from work that is merely blocked, and work that is finished from work that is finished but not yet reviewed. This page covers the states that earn their place and the nine state vocabulary Liquid Brain uses.

Why three columns are not enough

Classic kanban works because humans share context. Doing means someone is on it, and done means the team agrees it is done. With agents, each of those words hides a gap. An agent can be doing a task in the sense that a process is running, while nobody is checking its direction. An agent can report done in the sense that it stopped, while the work has not been looked at. Collapsing these into two columns invites the worst failure mode, where unreviewed output is counted as finished. Similarly, a single blocked column mixes two different situations: waiting on another task, and waiting on a person to make a choice. Only the second needs your attention, and only the first resolves itself. A board that distinguishes them lets you scan for what needs a human and ignore what does not. Work in progress limits help here as well, because unlimited parallel agents mostly produce unreviewed work.

States that gate what software may do

The test for a column is whether it changes what is allowed to happen. Backlog and planned mean not yet eligible for work. Active means an agent or person is on it. Blocked means nothing should start, and an automatic scheduler must skip it. Review means the output exists and a person must look. Completed means a person accepted it. Beyond the main line, three side states carry real meaning: a decision owed, which parks work until a person chooses; a hold, which pauses work deliberately without abandoning it; and declined, which records a choice not to do the work, so it is not reopened by accident. Crucially, these are available states, not a mandatory conveyor belt. Cards should jump to whichever state is true. Forcing every card through every column creates ceremony, and ceremony teaches people to click through without reading.

The nine states in Liquid Brain

Liquid Brain lets you track backlog, planned, active, blocked, review and completed work, with separate decision, hold and declined states [C2]. In the configured vocabulary these appear as BACKLOG, TODO, IN-PROGRESS, BLOCKED, VERIFY, DONE, DECISION-OWED, ON-HOLD and WONT-DO, and columns come from workspace configuration [C2]. They are options, not a required sequence [C2]. You can search by ID, title, domain, scope or artifact, filter by status and priority, switch between Kanban and List, and save project views [N01]. A project WIP limit can be set [N08]. The board is where you create cards and track them [C1a], and VERIFY means review is still needed, which is different from DONE and from a decision waiting for you. Because columns come from workspace configuration, your own board may name or arrange them a little differently [C2].

HOW LIQUID BRAIN DOES IT

Liquid Brain gives the board a nine state vocabulary so that status says something an agent and a human can both act on: you track backlog, planned, active, blocked, review and completed work, with decision, hold and declined kept separate ^[C2]. Those states are available options and not a forced sequence ^[C2]. You create cards and move them across the board ^[C1a], search by ID, title, domain, scope or artifact, filter by status and priority and switch to List ^[N01]. A project WIP limit keeps parallel work reviewable ^[N08]. Dependencies and blockers sit with the card ^[N03]. VERIFY marks work that still needs review, which is different from DONE ^[C2].

C2 Board vocabulary: Track backlog, planned, active, blocked, review and completed work, with separate decision, hold and declined states. Evidence: Implemented; nine-state test coverage in source
N01 Search, List and saved views: Find cards by ID, title, domain, scope or artifact; filter by status and priority, switch to List, and save project views. Evidence: Implemented; retained optional-management evidence and source
N08 WIP and weighted planning: Set a project WIP limit and opt into explicitly weighted completion using a chosen baseline. Evidence: Retained implementation/test evidence
C1a Board and card actions: Create project cards, edit their details, and track work through the board. Evidence: Owner-observed creation and moves; source-reviewed editing
N03 Planning and My Work: Keep assignee, reviewer, due-date, blocker and dependency details with project planning, and use My Work to focus attention. Evidence: Implemented; source and plan records

QUESTIONS

What columns should a kanban board for AI agents have?: At minimum: a not ready state, an active state, a blocked state, a review state for unreviewed agent output, a done state and a place for decisions waiting on a person. Liquid Brain uses nine states, from backlog through declined [C2].
Does every card have to pass through every state in Liquid Brain?: No. The states are available options, not a mandatory linear sequence [C2]. A card can sit wherever its real status is. VERIFY marks work that still needs review, and a decision waiting for a person is different from active implementation, so each state tells you who has to act next.
Can I limit how much work runs in parallel?: Yes. Liquid Brain lets you set a project WIP limit and opt into explicitly weighted completion using a chosen baseline [N08]. Only DONE earns completion weight, and missing weights or changed scope can make weighted progress unavailable, so treat the figure with care and read it alongside the board.

RELATED GUIDES

RELATED HARNESS PAGES

project cards vs chat threads Project cards vs chat threads: when a conversation with an AI agent is enough, and when a structured card record is the better home for the work.
human review of agent output Human review of agent output: what to inspect before accepting AI changes, and how recorded outcomes make the review honest.
multi-agent project coordination Multi-agent project coordination: keep Claude and Codex work on one project organized with shared cards, scope, authority and review.

LIQUID BRAIN AI HARNESS

COMING SOON

Verified against the October 17, 2026 development candidate (reviewed October 18, 2026)

READ THE GUIDES SEE THE CURRENT BUILD