Loops — Bounded Agent Loop Orchestration

A loop is a bounded feedback process (observe → choose → act → verify → record → repeat-or-stop) run over Maxim's existing offices/agents, with an explicit stopping condition and a named terminal state. This skill is the loop layer ON TOP of the office/command dispatch — it does not replace any command. Prior art: loop-library (the 31-loop catalog) and autonomy-loop (deterministic verification gates), both cited per ADR-007. Maxim adds the behavioral overlay: confidence tagging, framework citation, CSO auto-loop on regulated loops, and the no-fabrication rule.

---

Skill Purpose

Turn "do X repeatedly until it's right" from an open-ended autonomous run into a governed, finite loop with checkpoints and a defined stop. Maxim already owns the verbs (/mxm-build, /mxm-fix, /mxm-review, /mxm-test, /mxm-ship) and the offices; a loop orchestrates those against a stopping condition. The moat this skill protects: a Maxim loop never reports an error, a stall, or an exhausted budget as success — the confidence-tagging ethos (ADR-010) applied to iteration.

---

The Bounded-Loop Discipline (non-negotiable)

Every loop MUST declare, up front:

1. Trigger — what starts it (a command, a schedule, a backlog, a symptom).
2. Stopping condition — a rubric, threshold, benchmark, reviewer decision, or finite scenario set — never "until it looks good." If the operator gave no limit, use a no-progress stop; do NOT invent time/iteration/cost budgets.
3. Checkpoint — what is verified each iteration before continuing.
4. Terminal state — name it on exit: success · clean no-op (nothing to do) · blocked (dependency/permission) · approval-required (human gate) · exhausted (budget hit) · stagnated (no progress N rounds).

Hard rule: never report error, blocked, exhausted, or stagnated as success. Tag the exit state honestly (ADR-010). A loop that both produces and verifies its own work MUST run an independent verification pass (a separate reviewer invocation) — generation ≠ approval.

---

Activation & Dispatch

Trigger	Behavior
"run the <name> loop" / "iterate until X"	planner composes the loop, routes the per-iteration work to the owning office (ADR-017), enforces the stopping condition + checkpoints
A regulated/PII/data loop (e.g. production data cleanup)	CSO auto-loop fires — security-analyst gates every iteration; no PII copied into logs
A produce-and-verify loop	reviewer runs the independent verification pass each round

"iterate until coverage is 100%"  →  loops skill
   → planner composes: tester (act) + coverage checkpoint + reviewer (verify)
   → ratchet floor each round; stop at 100% (success) or no-progress (stagnated)
   → confidence tag + framework citation on exit

No command is required — this skill is invoked by intent, by /mxm-superpowers, or by an office command that recognizes a loop-shaped task.

---

The Loop Catalog (31 loops → Maxim surfaces)

Each loop orchestrates an existing Maxim office/agent. Status: ✅ already a Maxim command · 🔁 loop over an existing surface · ⭐ native gap-loop (defined below).

Engineering: 001 docs-sweep 🔁(proactive-watch+reviewer) · 002 architecture-satisfaction 🔁(build→review→test) · 003 perf-to-threshold ⭐ · 004 production-error-sweep ⭐ · 005 coverage-to-100 ⭐ · 007 logging-coverage ⭐ · 008 nightly-changelog 🔁(changelog-writer) · 011 test-suite-speed ⭐ · 016 ticket-to-PR-ready 🔁(/mxm-fix) · 019 adversarial-review 🔁(reviewer+pre-release-audit) · 020 scheduled-work-verification ⭐ · 027 builder-reviewer 🔁(/mxm-superpowers; see autonomy-loop) · 028 completion-contract 🔁(Coverage Matrix) · 031 recent-feedback-sweep ⭐

Evaluation: 009 quality-streak ⭐ · 010 full-product-eval 🔁(/mxm-health) · 023 champion-challenger ⭐ · 024 devils-advocate 🔁(reviewer/threat-modeler) · 029 versioned-experiment 🔁(experiment-tracker)

Operations: 012 repo-cleanup 🔁(/mxm-organize) · 013 stale-safe-batch-release 🔁(/mxm-ship) · 014 production-data-cleanup ⭐(CSO-gated) · 015 post-release-baseline ⭐ · 017 customer-deployment 🔁(customer-success-manager) · 025 fresh-clone ⭐ · 030 repo-maintainer-heartbeat 🔁(/mxm-portfolio+/mxm-tasks)

Content: 006 SEO/GEO ✅(/mxm-seo) · 018 product-podcast ✅(NotebookLM)

Design: 021 3D-render-benchmark (skipped — niche) · 022 frontend-reconstruction 🔁(ui-styling) · 026 thumbnail-iterate 🔁(banner-design)

---

Native Gap-Loops (⭐ — Maxim has no equivalent; defined here)

Ordered by priority. Each names its stop.

fresh-clone (025) — ⭐ top priority

In a disposable environment, follow ONLY documents/guides/GETTING_STARTED.md / README — no insider knowledge — until install + first-run works clean. Stop: clean install (success) OR a precise documented blocker (blocked). Targets Maxim's worst historical failure class: BUG-001..005 were all fresh-clone install failures (DEBUGGING_PLAYBOOK §1). This loop would have caught all five.

recent-feedback-sweep (031) — ⭐

Collect recent operator corrections (.mxm-operator-profile/ rejected patterns + .mxm-skills/agents-skill-gaps.log) → audit the whole project for the same pattern → fix each → record in the pattern registry. Stop: no new instances of the pattern (success) or no-progress (stagnated). Maps onto Maxim's pattern-registry discipline (BUG_TRACKER Recurring-Pattern Registry).

coverage-ratchet (005) — ⭐

Drive test coverage to a floor that can only ratchet UP (line AND branch, independently). A corrupt/missing baseline is an honest error, never a silent re-seed lower. Stop: floor met (success) or no viable test to add (exhausted). Ports the autonomy-loop coverage-ratchet logic.

production-error-sweep (004) — ⭐ (CSO-gated)

Group prod-log symptoms → separate actionable from noise → root-cause + smallest fix + verify, one PR each. Stop: clean if no actionable errors (clean no-op) or all fixed (success). security-analyst gates: never copy credentials/PII into logs or output.

champion-challenger (023) — ⭐ strategic

Promote a prompt/policy/framework change only if it beats the incumbent on a frozen holdout set without weakening a must-pass check. Keep the incumbent on uncertainty. Stop: target met / budget / no-progress. On-ramp to the PaaS data flywheel — the only loop that makes Maxim self-improving rather than static.

perf-to-threshold (003) · logging-coverage (007) · quality-streak (009) · test-suite-speed (011) · data-cleanup (014, CSO-gated) · post-release-baseline (015)

Defined in references/loop-catalog.md (each: trigger · stop · checkpoint · owning agent).

---

Behavioral Overlay (applied to every loop)

• Confidence tag on every iteration's output and on the terminal state (ADR-010).
• Framework citation — name the behavioral/CS principle the loop embodies (ADR-007): bounded loop (feedback control), independent verification (separation of duties), champion-challenger (Goodhart-resistance + holdout).
• CSO auto-loop — any loop touching PII / regulated data / production data routes security-analyst into every iteration; cannot be bypassed.
• No-fabrication — any rate/metric a loop reports carries its sample size (e.g. "12/30 passing"), never a bare claim. (autonomy-loop's honesty rule.)

---

References

• loop-library — Forward-Future/loop-library (31-loop catalog; the stopping-condition discipline). Prior art, cited per ADR-007.
• autonomy-loop — inferencegod/autonomy-loop (deterministic fail-closed gates: coverage-ratchet, mutation-as-gate, money-path floor). Reference for the verification loops.
• ADR-002 (executable contracts) · ADR-007 (framework citation) · ADR-010 (confidence) · ADR-017 (office dispatch).
• Full adoption rationale + the 31-loop mapping: documents/architecture/MAXIM_LOOP_ADOPTION_PLAN.md.