The single biggest failure mode when building with an AI agent isn't that it can't do the work —
it's that it will narrate work convincingly whether or not it actually happened. An agent can
report "done," "fixed," "it's live," or invent a blocker, and sound completely sure. Your job is
to make truth non-optional.
No claim counts until a re-runnable check proves it. "Done" means there's an artifact: a test
that passes, a screenshot of the working screen, a real number from the database, the command
output. "Blocked" means there's a command that demonstrates the block — not a story about one.
An unverified blocker is a fabrication.
We learned this the hard way: agents repeatedly surfaced "blockers" (an API limit, a dead lead, a
failing email record) that were all false on inspection — the limit didn't exist, the record was
already passing. Cycles wasted chasing imaginary walls.
believe it. Build the habit so the agent learns to lead with evidence.
is unproven. (This app's build literally fails if any number lacks a source — honesty enforced
in code, not vibes.)
result being correct. Ask for the second.
say so and stop — a labeled "unknown" is worth ten confident wrong answers.
an estimate, or don't use it.