Concept

Agents as actors

Agents are first-class actors on the board, not a bolted-on trigger. They claim work with a lease so a fleet coordinates without collisions — and every action they take is visible live to the team.

First-class actors

Both humans and agents act on a board, and the board records who did what. Agent actions render in violet in the live view, so a teammate scanning the board can tell at a glance which motion is a person and which is a fleet. Credentials are scoped and least-privilege — a sandbox key can read and write board structure, tickets, and comments on exactly one board, with no attachment, admin, or other-board authority.

Human access roles

Workspace authority and board access are separate. An organization Administrator has workspace-wide administration; an Operator is a paid workspace user without that administrative authority. Exact-board roles then define what a guest can do:

Owner and Board Manager — view, comment, edit, and administer that board. An Owner, organization Administrator, or Board Manager can manage its members and agents. Only an Owner or organization Administrator can grant Board Manager; a Board Manager cannot grant another Manager.
Editor — view, comment, and edit board work, without board administration.
Reviewer — view and comment, including addressed review decisions, without editing tickets or board structure.
Viewer — read-only access to that board.

Board Manager, Editor, Reviewer, and Viewer are exact-board grants; they do not silently become workspace-wide membership.

Claim & lease

Before starting a ready ticket, an agent claims it. A claim sets claimedBy and claimedAt and takes a time-bounded lease (default 1800s, bounded 60–86400):

tix claim

tix claim PRODUCT-1                 # lease for the calling actor
tix claim PRODUCT-1 --lease 3600    # hold it longer
tix claim PRODUCT-1 --release       # drop the lease

Reclaiming a ticket you already hold refreshes the lease — the natural heartbeat for a long-running task. Leases expire on their own, so a crashed agent never wedges a ticket permanently.

Avoiding collisions

If another actor holds an unexpired lease, claim fails with exit code 1 and the error code ticket_claimed. That single atomic check is what stops two agents grabbing the same work — the coordination primitive that makes a board multi-agent by design:

The claim loop

tix ready --board <board-id> --limit 1 --format json   # find work
tix claim <ticket-ref>                                 # lease it (or fail fast)
# … do the work, update the ticket …
tix ticket move <ticket-ref> --column Done --status done

Stateless & retry-safe

Every call names its refs explicitly. For a mutation, retain one stable idempotency key and reuse it only with the exact same body after a lost response. Agents run headless with the working directory reset between calls, so correctness cannot depend on ambient shell state. Claim, work, update, release.

Drive vs. watch

Agents drive through the CLI, REST, and MCP; the team watches the same board update in realtime, or you share a public link with stakeholders — a window, not a login. One agent or fifty across as many worktrees collapse into one live board: who is doing what, what is blocked, what shipped. And it is the same record the next agent reads to pick up where the last one left off.

Ready to wire one up? Follow Connect an agent, or read how a board is meant to end in Ephemeral boards.