Journal
What we're learning
Notes from building with agents: context engineering, agent harnesses, and what our product factory gets wrong.
Local proof is the gate; continuous integration is the backstop
Our merge rule stated plainly, with local verification as the real gate, red checks as a hard stop, and absent checks handled as a recorded coverage gap.
Read the entryContext selection is a policy that can be measured
Five papers treat the context an agent receives as a selection policy that can be measured: what to keep, what to render, what to disclose, and which records a decision depends on.
Our delivery train does not wait for review
How our agent delivery train keeps work moving while code review runs: parallel lanes on separate files, review as a queue, and a failed review as a new task rather than a blocked lane.
Defending the execution boundary for agent actions
Three papers defend the point where an agent's internal state becomes an external action instead of chasing attack patterns, and their numbers show what that means for our own authorisation checks.
A timeout was reported as a parsing error
A build agent killed by its time budget was reported as producing unparseable output, sending self-repair after the wrong problem twice; the fix names the true failure and gives the reporting component its first tests.
The harness is the lever, until it is not
Two papers show where harness evolution pays, where it reaches its ceiling, and why splitting failures into process and content changes the fix.