How outerloop compares

outerloop is not a replacement for code review or CI. It is the governance layer that makes agent-assisted work answerable after the fact.

vs. GitHub PR reviews

	PR review	outerloop
When	At merge time	At agent-run completion and again at ship
Artifact	Comments, approvals	EvidencePackage + Verdict + Ledger chain
Rationale	Often implicit or scattered	Mandatory, structured rationale
Agent runs	Usually invisible	First-class evidence from harness output
Retro	Dig through threads	ledger why <sha> reconstructs full chain

Use both: PR review for code quality; outerloop for agent governance and answerability.

vs. CODEOWNERS / branch protection

Branch protection enforces who may merge. outerloop enforces why a change was accepted when agents did most of the work.

• CODEOWNERS: static file ownership
• outerloop: dynamic evidence, risk scoring, taste rules, verdict rationale

vs. conventional CI (tests, lint, SAST)

CI answers "did checks pass?" outerloop answers "what did the agent do, who decided it was safe enough, and can we explain that six months later?"

# CI gate
pnpm test && pnpm lint

# outerloop gate (composable)
outerloop evidence package --run-id "$RUN_ID"
outerloop policy evaluate "$EVIDENCE_ID"
outerloop verdict issue --evidence-id "$EVIDENCE_ID" --decision ship --rationale "..."

vs. loop-engineering

loop-engineering designs and runs inner loops (patterns, state, verifiers). outerloop owns the outer loop (evidence, verdict, ledger).

They compose: loop-engineering produces runs; outerloop packages evidence and records human verdicts.

vs. "just document in the PR description"

PR descriptions are free-form and optional. outerloop makes rationale required, typed, and queryable — linked to structured evidence and indexed in a ledger.

When outerloop is the wrong tool

• One-off scripts with no agent involvement (use normal git + PR).
• Teams that only need test gates (CI is enough).
• Replacing security review for regulated systems (outerloop complements, does not replace specialists).

When outerloop fits

• Agentic coding workflows (Cursor, Claude Code, custom harnesses).
• High-agency inner loops where humans must stay accountable.
• Teams scaling multiple agent loops and feeling orchestration tax.
• Audits asking "why did this ship?" months later.