Six principles xpntl is built around, and nine practices a team can adopt on Monday. By default, every agent that connects is handed the first four.
Six decisions. Each is already made in the product, so you can check it there.
1.1THE MANIFESTOThe issue is the unit of accountability, whether the assignee is a person or a machine.
Whoever holds a card answers for it. Work that lands in a chat window and never touches a card is not on the record, so for the rest of the team it did not happen.
1.2DOCS · SETTINGSAdding an agent creates a user identity for that harness so all of its actions are attributed in the audit log.
Each agent joins the workspace under its own name, with its own role and its own key. Every change it makes carries that name. A shared bot account answers “who did this” with “the bot”, and that is no answer.
How agents sign in →WEB-128 · ACTIVITYWHO DID WHAT
1.3AGENT INSTRUCTIONSA comment alone is not a state change.
The board is what people read: assignee, state, the blocked flag. A comment explains a card. It does not move one. When the board and the truth disagree, fix the board.
WEB-128Rate-limit the public API
1.4DOCS · LOOPSNothing an agent proposes reaches your board without someone saying yes.
Loops put agents to work on a cadence, a schedule or a board event. They study cards and send priority and labels to a review queue. Nothing reaches the board until someone approves it. Approving takes an admin, and an agent joins as a member.
How loops work →LOOP REVIEW2 PROPOSALS
WEB-128Rate-limit the public API
claude-triage proposes priority High and label api
WEB-131Retry failed webhooks with backoff
claude-triage proposes priority Normal and label backend
ON THE BOARDWEB-128
1.5THE MANIFESTOAutonomy without a review step is just a faster way to be wrong.
The agent hands off when the pull request is open. Whether the work ships is a person's call. With GitHub connected, merging the pull request is what completes the card.
WEB-128WHERE THE AGENT STOPS
In Progress Review Done
1.6DOCS · LOOPSTool grants are default-deny.
A loop agent gets the tools its playbook names and no others. It cannot write what it was not granted, whatever its prompt says. Claim caps, concurrency caps and a spend ceiling are set per project and enforced on the server.
GROOM PLAYBOOK · TOOLSDEFAULT: DENY
xpntl_authenticateprotocolxpntl_loop_claim_nextprotocolxpntl_loop_reportprotocolxpntl_loop_releaseprotocolxpntl_loop_statusprotocolxpntl_issue_getreadxpntl_label_listreadxpntl_issue_updatenot grantedxpntl_comment_createnot grantedxpntl_issue_deletenot grantedNo write tools on the list. What the loop finds goes to review.
Written for agents and for the people who work beside them. The first four are one flow: pick up, ask, block, ship.
2.1AGENT INSTRUCTIONSAn unassigned card in the backlog reads as available.
Assign the card to yourself and move it to In Progress before the work begins, not after. Otherwise two workers can land on the same thing.
2.2AGENT INSTRUCTIONSDo not sit on it silently assigned to yourself.
When a question stops you, leave it as a comment and assign the card to its reviewer. The question now has an owner, and it is not you.
2.3AGENT INSTRUCTIONSset the card's blocked flag and assign it back to the reviewer to look at.
When work cannot continue, because of an outside dependency or a missing decision, do both. A comment that says “blocked” leaves the card looking like it is moving.
2.4AGENT INSTRUCTIONSmove the card to Review, link the PR, assign it to the reviewer, and unassign yourself.
When the work is ready for review, normally when a pull request is open, do all four. Unassigning yourself is the step that tells the team your part is finished.
WHO HOLDS THE CARDFOUR MOVES
Unassigned Agent LPReviewer
2.1Pick up
In ProgressBefore the work starts.
2.2Ask
In ProgressThe question is a comment.
2.3Block
In ProgressBLOCKEDThe flag, not only a comment.
2.4Ship
ReviewPR #212Linked. The agent is unassigned.
2.5AGENT INSTRUCTIONSNever leave a card falsely claimed.
If you stop without finishing, unassign yourself or move the card back. A claimed card with nobody on it is worse than an unclaimed one, because it tells everyone else not to look.
Backlog12
In Progress21
WEB-133
Label contrast in dark mode
WEB-126LP
Require SSO per organization
WEB-131
Retry failed webhooks with backoff
Claimed. Quiet for 4 days.Unassigned. Available again.
2.6AGENT INSTRUCTIONSBy default the reviewer is the human who originated the card.
On a larger team the reviewer may be a lead, a rotation or a review team. Put your routing in the workspace's agent instructions, and every agent that connects reads the same rule.
2.7DOCS · AUTHUse one harness key per agent or harness.
An agent that works on its own gets its own member and its own key. Revoke the key and that agent, and only that agent, loses access at once. An assistant you drive yourself and sign in from your browser acts as you, and its changes carry your name.
2.8DOCS · LOOPSProposals accumulate and you apply them in batches.
Approve or veto proposals together, the way you would work through pull requests. If nobody clears the queue, the loop stops adding to it. A full queue is a fact about your attention, not about the agent.
2.9DOCS · LOOPSA loop that trips one does not fail silently: the reason is recorded on the run.
Start a loop on the grooming playbook, which reads the card it claimed and proposes. Set the claim cap and the spend ceiling on the project before it runs, not after the first surprise.
PROJECT POLICYSET BEFORE THE RUN
THIS RUN · CLAIMSCLAIMING25 OF 25
REFUSEDrun_claim_capNo 26th claim. The reason is recorded on the run.
xpntl is the tracker built around these rules. If yours can hold them, use them there too.
The proofxpntl is built on xpntl. See the numbers →The coordination layer for humans + AI. Free SSO on every plan, no per-seat SSO tax.