BBL/FIELD MANUAL — FREE DOWNLOAD

The AI Agent Operating Playbook

40 field rules for running AI agents that actually ship — mined from hundreds of hours of real production sessions. Not prompt tips. The judgment calls that separate “looked done” from “actually worked”.

Every rule below was paid for with a real failure. Read the first section free.

Trust nothing stale

Re-check every external blocker fresh before continuing a plan. Credit balances, API tokens, approval gates. Yesterday’s “it works” is a rumour, not a fact.
Before extending or “fixing” a system, verify it actually exists. Repo, disk, host, DNS. Ghost projects get flagged, not silently rebuilt — rebuilding twice is how months disappear.
“I thought we built X” usually means “deliberately deferred”. Check the old plan before assuming failure, and never let a deferral live only inside a document — make it a visible follow-up.
Distrust caches. DNS answers, media players showing old files, dashboards lagging reality. Verify against the authoritative source.
Small tool lies compound. A counter counting the wrong thing, a shell resetting its directory — verify the measurement before trusting the number.

Root-cause upstream

Check the suspected cause first — with data — then follow the evidence. When it’s ruled out, say so explicitly so everyone’s mental model gets corrected too.
The bad output you can see is days downstream of the real failure. Read the failure logs of prior runs before staring at the artifact. Ask: how long has this actually been broken?
After fixing a pipeline, repair everything it produced while broken. Not just the reported cases. Sweep the whole window, and re-check earlier repairs for what they missed.
Name the fixes you rejected and why. A fix chosen without alternatives considered is a guess wearing a suit.
Never fix a race condition with delays or random jitter. Jitter turns a daily, debuggable failure into a rare, untraceable one. Serialize or lock, then add retries as backstop.

UNLOCK ALL 7 SECTIONS / 40 RULESFREE