Agents in the pipeline: cheatsheet

Verified automation flags

Flag	What it does
-p, --print	Run one prompt non-interactively; read stdin, print, exit
--output-format	text (default), json (result, session ID, metadata, cost), stream-json (event stream)
--json-schema	Structured output conforming to a JSON Schema (with JSON output, print mode)
--allowedTools	Pre-approve tools; there is no prompt to answer in a pipeline
--permission-mode	Session baseline; dontAsk suits locked-down CI runs
--max-turns	Cap agentic turns; exits with an error at the cap (print mode)
--max-budget-usd	Dollar ceiling for a run (print mode)
--append-system-prompt-file	Pass explicit instructions from a file
--bare	Skip auto-discovery of memory, hooks, plugins; same result on every machine
--continue, --resume	Share context across runs; never between generator and reviewer

Spine of the lesson: an unattended agent is the purest test of your architecture.

Two jobs, two budgets

	Nightly batch job	Blocking pre-merge check
Who waits	Nobody	A developer
Scope	Wide (a whole module)	Narrow (the diff only)
Speed	Slow is fine	Fast and predictable
On failure	Log, retry tomorrow	Fail fast and visibly
Watch	Dollars per run	Minutes per run

Generator vs reviewer

Generator	Reviewer
Carries its own reasoning and assumptions	Fresh context, fresh process
Reads what it meant to write	Sees only the artifact and the standards
Goes easy on itself	Catches more
Its pull requests still get reviewed	Never continued from the generator’s session

The dedupe recipe

1. Fetch the bot’s prior comments on the pull request.
2. Include them in the prompt as prior findings.
3. Instruct: report only new or still-unaddressed issues; never repeat resolved ones.
4. Post only the delta. State lives in the pull request, not the agent.

Writing review criteria

Do	Do not
List explicit finding types	Say “be conservative” or “only what matters”
Give each type a concrete example	Make the model guess the threshold
Name what must never be flagged	Leave style and naming to opinion
Grant permission to find nothing	Reward findings invented for thoroughness

Pitfalls

Pitfall	Correction
Generator reviews its own output	Separate invocation, no shared context
Criteria as moods	Examples pin the threshold
Reposting findings each push	Carry prior findings into context
One budget for every job	Batch and blocking need opposite treatment
Standards pasted into workflow files	Standards live in checked-in project memory

One-liners

Line	Meaning
Nobody is there to catch it, so the design must	Every decision made in advance, in files
A false-positive flood spends the bot’s only currency	The team’s willingness to read it
The CI agent is the newest teammate	The repository does the onboarding
A slow blocking check gets deleted, not tolerated	Latency is architecture too