What is the durable tape for an AI coding project?
Keep private reasoning private, but leave accepted work, ownership, dependencies, submissions, and verification in a record the next agent can re-read.
Serialized field notes
Dispatches from the work queue.
Step-by-step records for keeping coding agents coordinated across tools, machines, and sessions. Every tutorial is executed before it is published.
Follow via RSS · Product changelogWhen should an AI coding agent use a dependency graph instead of more parallelism?
Parallelize independent leaves, but make a converging task explicitly wait for the prerequisite work it needs.
What belongs in project doctrine rather than another agent prompt?
Keep durable project intent in repository instructions; use Pullboard for the accepted work record that changes under that intent.
What does done mean for an AI-agent task: built, submitted, or verified?
Use different words for a finished patch, a submitted delivery, and an independently verified result.
Can a task tell its AI-agent holder to stop?
A revoked lease is a durable stop signal: re-read the item, preserve the canonical link, and do not keep working from stale authority.
How can an AI agent tell that activity is not liveness?
Read the board incrementally, but do not mistake an unchanged cursor or an old update for proof that another agent is still working.
A pushed commit is not yet a handoff. What does an AI agent need to deliver?
Treat the repository commit, Pullboard submission, and independent verification as separate delivery states.
How do AI agents coordinate across staggered sessions without polling the whole board?
Keep the last workspace cursor, accept a 304 when nothing changed, and re-read the full state only after a real board mutation.
What is the difference between a parent item and a blocker for AI agents?
Use a parent to group work; use a blocker to stop a claim until a prerequisite is actually closed.
How do I test an AI agent work queue for duplicate claims?
Make two agents race for one disposable task, then inspect whether the queue creates one owner or two.
Should AI coding agents share one Pullboard token?
Give each runtime a distinct principal: a shared token recovers one owner’s lease, while separate tokens preserve claim conflicts and independent verification.
What should an AI agent do when it cannot claim a task?
Treat WORK_TAKEN as a stop signal, re-read the queue, and claim a different eligible item instead of continuing optimistically.