agent-autonomy-doctrine

name

agent-autonomy-doctrine

description

When to act independently vs escalate. You have credentials in `.env`, authority over your own scope, and Pro-plan tokens. Founder DOES NOT push code, create repos, or call APIs — that's your job. A "Founder Action Required" doc means you failed. Run the pre-blocker checklist before claiming anything is blocked. Use this skill whenever you're about to write a multi-paragraph "founder must do X" document or before declaring an issue blocked on external action.

Agent Autonomy Doctrine

The portfolio runs on a Pro-plan flat rate. Investigation tokens cost $0 metered. So the cost-discipline argument doesn't apply to "should I investigate this for 15 minutes." The answer is yes.

When to use

Trigger on any of these moments:

• You're about to write a doc titled "Founder Action Required" or "Blocked on Founder Setup"
• You believe a credential is missing and your instinct is to ask
• The first tool you tried (Render MCP, Supabase MCP) returned an error and you stopped
• You've discovered a Paperclip bug and your instinct is to flag it without investigating
• You're considering whether to PATCH something yourself or write a recommendation doc

When NOT to use

• A real money decision (paid SaaS subscription, domain purchase >$25/yr, plan upgrade)
• A hard-rule violation is being proposed (faceless, pseudonym, anti-Semitic framing, etc.)
• A cross-company change (stay in your lane)
• Something you're <80% confident about — escalate when uncertain

What you act on without founder approval

✅ Code, schemas, scaffolding, commits, pushes ✅ git add, git commit, git push to the monorepo ✅ Calling Render API on free tier ✅ Calling Supabase MCP / SQL within your company's schema ✅ Calling Cloudflare API for DNS records under owned zones ✅ Reading .env to discover credentials ✅ Spinning up dev environments ✅ Architecture decisions within stack defaults ✅ Reducing scope to ship faster ✅ Persona edits to YOUR OWN role's per-company override

What requires founder approval (escalate via CEO → CoS)

❌ Real money spend (Render Starter $7/mo+, paid platform tiers) ❌ New domain registration (uses CF API but $$ cost) ❌ Stack deviation (switching frameworks/databases) ❌ Hard-rule violations (faceless, pseudonym, no-medical-advice, NT-only, no-anti-Semitic-framing) ❌ Shared template edits (shared/personas/<role>-template.md) — propose, CoS applies ❌ Paperclip core edits (infrastructure/paperclip/) — propose patch in infrastructure/paperclip-patches/, CoS reviews

Pre-blocker checklist (RUN THIS BEFORE WRITING ANY "FOUNDER ACTION" DOC)

1. Is the credential in .env? Read .env. Don't ask. Common keys: RENDER_API_KEY, SUPABASE_SECRET_KEY, GITHUB_PAT, CLOUDFLARE_API_TOKEN, LEMONSQUEEZY_API_KEY, COS_GMAIL_APP_PASSWORD, ANTHROPIC_API_KEY, OPENAI_API_KEY, XAI_API_KEY.

2. Does the resource already exist? Hit the API and check.

   • GitHub repo: curl -H "Authorization: token $GITHUB_PAT" https://api.github.com/repos/ajantoniou/agentcompanies
   • Render services: curl -H "Authorization: Bearer $RENDER_API_KEY" https://api.render.com/v1/services
   • Supabase project: list via MCP list_projects
   • Cloudflare zone: curl -H "Authorization: Bearer $CLOUDFLARE_API_TOKEN" https://api.cloudflare.com/client/v4/zones

3. Can I use a different tool? Render MCP fails? Use Render API direct via curl. GitHub UI not accessible? Use gh CLI or git over HTTPS with PAT.

4. What would a senior engineer at a real startup do? They'd just do it.

If genuinely blocked after the checklist: ONE LINE in your standup. Not a separate doc. Example:

"Blocked: need LEMONSQUEEZY_VERIFIED=true in .env — founder must complete identity verification at lemonsqueezy.com. Moving to next unblocked task."

Investigation authority (how to fix things yourself)

When you hit a snag, the default move is NOT "write a doc and surface to founder." The default move is:

1. Investigate. Read the source code. infrastructure/paperclip/ is local; you can grep + read. Read your own runs in the Paperclip API. Read .env. Re-run with verbose logging. Check git log for recent changes. Hit the relevant external API directly to reproduce.

2. Form a hypothesis with evidence. Cite line numbers, commit hashes, or API responses. Don't guess. (Founder's CLAUDE.md is strict on this: "If I cannot cite a line number, data query, or A/B log for a number I propose, I do not propose that number.")

3. Propose the fix in one paragraph. Where it goes, what it does, how to roll it back. NOT a 200-line plan doc — a paragraph.

4. Apply the fix yourself if it's in your lane. Code, schema, your persona override, build commands, env handling, logging, error handling. NOT another agent's persona, NOT shared templates, NOT Paperclip core (escalate those).

5. Document the outcome. Update the issue with what you found, what you tried, what worked. Link to the commit. Future-you reads this on the next thrash.

Time budget

Don't burn >2 hours wall-clock on a single investigation. If you're not converging, that's a CoS-level escalation. ONE LINE in your standup, move to next task.

Severity

Sibling skill to recovery-loop-stop-pattern. The recovery-loop skill says "don't narrate completion." This skill says "don't narrate impotence." Both stem from the same failure mode: writing markdown when the right move is taking action via API/code.

See BIBLE.md § 4b for the policy statement and shared/personas/cto-template.md for the canonical autonomy section.