engine

agent-markdowns

A Claude Code plugin marketplace for autonomous, self-verifying agent work — the agent-loops toolkit.

The toolkit: one loop engine, four modes

The loop is always the same — produce an artifact → verify it against something that can fail → adversarially attack it → simplify → iterate to a machine-checkable done. Only the bindings change per mode. Each mode ships as its own plugin, so you install à la carte:

Plugin	Mode	What it's for
agent-loops	foundation	The operating manual every mode rests on — the playbook + the goal-template. Install this first.
engine	build	Build a feature/fix autonomously and self-verify until green, stopping at a committed slice.
review	review	Audits, security sweeps, deep reviews — findings as the product (no repro = no finding).
planning	plan	Design a spec/plan before any code; emit a machine-checkable plan build mode consumes.
infra	infra	Migrations / deploys / backfills — the guardrail harness (parity, canary, rollback) comes first.

The mode plugins are self-contained but richer with agent-loops installed alongside — their skills reference its playbook for the full method.

Layout

plugins/
├── agent-loops/   skills/engineering-agent-loops/
│                    SKILL.md
│                    references/agent-loops-playbook.md   # the accumulated judgment
│                    templates/goal-template.md           # build-goal stations
├── engine/        skills/engine/      SKILL.md + templates/engine-template.md
├── review/        skills/review/      SKILL.md + templates/review-template.md
├── planning/      skills/planning/    SKILL.md + templates/planning-template.md
└── infra/         skills/infra/       SKILL.md + templates/infra-template.md

Three layers, three lifespans: a skill routes (per-trigger), the playbook accumulates (per-repo, grows over time), and your task ledger stages (per-task) — the layering that survives context compaction. The playbook is the distillation of real autonomous runs — among them a 7-phase semantic-search + RAG build (~4,900 lines across schema, indexing, embeddings, retrieval, backfill, frontend, monitoring) whose adversarial verify panel caught a real bug every phase, including two security issues no compiler, lint, or test could see — plus an overnight-lane night-shift retro and a multi-PR harvest off a moved base branch.

Install

# Add this marketplace
/plugin marketplace add wilrf/agent-markdowns

# Install the foundation + whichever modes you want
/plugin install agent-loops@agent-markdowns
/plugin install engine@agent-markdowns
/plugin install review@agent-markdowns      # optional
/plugin install planning@agent-markdowns    # optional
/plugin install infra@agent-markdowns       # optional

Use

/engine <your task>          # build mode — directly invocable

…or just describe the work and the right mode triggers on its own: "audit this for security" → review, "plan the X feature first, don't code" → planning, "migrate this table safely" → infra. /engine and the engineering-agent-loops skill are directly invocable; the review / planning / infra skills fire on their trigger phrases.

License

MIT