The supervised AI operating model for software teams.
Three supervised peer loops across the full SDLC — mechanical gates the agent cannot bypass, cold-read specialist reviewers, and human checkpoints it cannot self-certify past.
Reversible. One command to try, one to remove.
- 3supervised loops
- 7adapters
- 1pip install
An unattended loop makes unattended mistakes.
Agent coding tools have moved the leverage from the prompt to the loop. But a loop running unattended is also a loop making mistakes unattended — and it will self-certify its way out of every check you give it. The answer is not a tighter prompt. It is a supervised loop with mechanical gates the agent cannot bypass.
Three supervised peer loops. One handoff chain.
Discovery Loop
product-engineering · discovery-leadRaw idea → ratified brief. Five candidate shapes explored in parallel, collapsed through product, UX, architecture, and safety lenses.
Human gate. G3 — you approve a ratified decision brief, not a validated solution.
Read the discovery journeyBuild Loop
core · work-loopSpec → shipped code. Lint, typecheck, and tests must pass. Three specialist reviewers each read the diff cold, in a fresh session — adversarial by design.
Human gate. G4 — you merge only when adversarial review is clean and all gates pass.
Read the build journeyRelease Loop
release-engineering · release-leadBuilt → production. Autonomous e2e convergence on ephemeral environments. Deployed findings feed back to the build loop automatically.
Human gate. G5 — prod ship always surfaces to a human. Always.
Read the release journey
Every loop surfaces to you at exactly the right moment.
Not as a workaround — as a design principle. The agent runs the loop; you control the gates. Here is what each handoff asks of you.
- G0Discovery
Go / No-go
Is this idea worth exploring?
- G1.5Discovery
Shape sign-off
Does the chosen shape fit the product strategy?
- G2Discovery
Brief ratification
Is the decision brief complete enough to build from?
- G3Discovery → Build
Handoff to build
Is this brief specific enough to spec?
- PlanBuild
Plan approval
Is the plan's trio correct? Do the risk triggers fire?
- G4Build
PR merge
Is adversarial review clean? Does the spec match the implementation?
- G5Release
Prod ship
Is this safe to put in front of users?
One install. Every major agent.
| Agent | Skills | Subagents | Hooks | Commands |
|---|---|---|---|---|
| Claude Code | Yes | Yes | Yes | Yes |
| Codex | Yes | Yes | Yes | No |
| Cursor | Yes | Yes | Yes | No |
| Copilot | Yes | Yes | Yes | No |
| Gemini CLI | Yes | Yes | Yes | No |
| Kiro IDE | Yes | Yes | Yes | No |
| Kiro CLI | Yes | Yes | Yes | No |
Switch adapters with one flag. Your skills, subagents, and hooks project into the layout each agent expects.
Start in one command.
pip install agentbundleagentbundle install --pack coreagentbundle install --pack coreagentbundle install --pack product-engineering --scope useragentbundle install --profile inceptionagentbundle install --profile solution-architectOne command lands the loop. Any agent that reads a skill file inherits it automatically.
The catalogue.
A curated library of packs — each shaped through practitioner research and RFC governance. Start with the loops.
Core
repoThe build loop. work-loop, new-spec, bug-fix, four specialist reviewers, hooks. Install this even if you install nothing else.
Read the journeyagentbundle install --pack coreProduct Engineering
userThe discovery loop. discovery-loop, frame-intent, de-risk-intent, decompose-intent, voice-and-microcopy, align-value-stream.
Read the journeyagentbundle install --pack product-engineering --scope userRelease Engineering
repoThe release loop. release-loop, release-lead. Autonomous e2e convergence. Hard-depends on core.
Read the journeyagentbundle install --pack release-engineering
See all 17 packs
Install once, works across all your repos user
Product Strategy
The strategy seat — SWOT, Porter, PESTLE, OKR cascade, PRFAQ, plus UX and content strategy.
Desk Research
Evidence-grounded research — surveys, source maps, competing-hypothesis analysis.
Experience Design
The full design thread — journey to realization, WCAG-aware quality floor.
Architect
System design — design docs, diagrams, architecture review.
Contracts
API-first design — OpenAPI 3.1 and AsyncAPI contracts.
Converters
Document conversion — PDF/DOCX/PPTX ↔ Markdown, HTML, Office, Mermaid.
Credential Brokers
Credential resolution — env → OS keyring → dotfile. Cleartext never reaches the model.
Atlassian
Jira and Confluence from the agent — flow and DORA metrics, SSO-cookie auth.
Figma
Read and render Figma designs — files, nodes, variables, FigJam → Mermaid.
Install per project repo
IaC (Terraform)
Plain-language intent → governed, cloud-agnostic Terraform. Human-gated; stops at plan.
Governance Extras
Decision trail — new-rfc, new-adr, update-conventions for long-lived repos.
User Guide (Diátaxis)
Docs scaffold — Diátaxis skeleton with new-guide.
Monorepo Extras
Package scaffolding — new-package, example package template.
Catalogue Curation
Grow the catalogue — assimilate primitives, survey repos, propose packs, export forks.
Make it your team's operating model.
Adopt the catalogue as-is, or fork it as your own. Write your house conventions and review standards into core, add skills for your stack, and ship one catalogue every engineer installs in a single line — the loop, the reviewers, and the standards come out identical on every machine and in every agent.