binky-control/AutonomyPolicy.md

Autonomy Policy

Status: draft v0.1 — 2026-07-16. Governs what agents may do without the founder. In force once ratified. The operating rule: "ask Bernd" is never the default.

The rule

If the founder is unavailable, the system must either continue safely, prepare the next decision, or explicitly defer with evidence. It must not silently idle.

Valid states: proceeding | prepared for review | deferred by policy. Invalid state: waiting because unsure. Uncertainty triggers research, comparison, preparation, or risk classification — not paralysis.

Autonomy lanes

Lane	Name	Agent may	Human needed?
Green	Safe autonomous	Research, summarize, classify, draft, organize, test, prepare	No
Blue	Reversible execution	Internal docs, backlog updates, local scripts, reports, local commits	No, but logged
Yellow	Deferred approval	Prepare external emails, invoices, offers, releases, payments	Yes, async via decision queue
Orange	Time-window approval	Customer/legal/accounting/banking actions needing business hours	Yes, on office-hour days
Red	Human-only	Signing, strategic commitments, bank changes, sensitive legal/HR/security	Always

Optimization target: maximize Green+Blue throughput, queue Yellow, batch Orange into office-hour days, protect Red from automation.

Default policies

• Research: if uncertain, collect evidence and produce a comparison. Ask only if ambiguity changes risk or external commitment.
• Documentation: agents may create/improve internal markdown freely; must preserve provenance and mark assumptions.
• Residuals (finished workplans): before marking a workplan finished, hand off actionable leftovers as live work records (intake with origin: residual + origin_ref: <WP-id>, or a next workplan / decision / engagement). Green/Blue residual capture is normal autonomous work. Do not park residuals only in prose or SCOPE.md. Fleet canon: work-record-types_v0.1.md § Residuals.
• Email: draft freely; send only pre-approved low-risk routine categories. Pricing, commitments, complaints, contracts, legal → decision queue.
• Spending: recommend freely; never spend outside approved recurring budgets or explicit thresholds.
• Product: refine specs, use cases, architecture, backlog freely; strategic direction is never final without founder approval.
• Code: branches, tests, prototypes, docs, local commits freely; no production deploys without policy-approved checks.

Approval packages

When an agent hits a lane boundary it prepares a complete decision package, not a bare question:

decision: ""
recommended_action: ""
reason: ""
evidence: []
risk_level: yellow
reversibility: low | medium | high
deadline: ""
attention_cost: "15 min"
fallback_if_no_approval: ""
options: [approve, reject, revise, defer]

Founder attention budget

daily_attention_budget:
  available_minutes: 90        # typical of the 1–3 h window
  max_decisions_per_day: 7
  max_deep_topics_per_day: 1
  interaction_style: [concise briefs, approve/reject options, evidence links]

Founder availability is first-class state

founder_availability:
  state: available | evening_window | unavailable | office_hour_day
  allowed_escalation_when_unavailable:
    - critical security incident
    - payment failure
    - legal deadline

On return from absence, agents produce a Return Brief: what changed, what progressed, what waits, what went stale, what risks increased, recommended decisions for today.

Initial autonomy thresholds

Area	Autonomous?	Limit
Internal docs	Yes	Preserve previous versions
Research	Yes	Cite/store sources
Backlog refinement	Yes	Propose, don't delete major items
Code prototypes	Yes	Branch only, no production deploy
Customer emails	Draft only	Send after approval
Social/publishing	Draft only	Publish after approval
Payments	No	Prepare only
Contracts	No	Analyze/draft only
Security changes	Mostly no	Prepare/test only
Cloud infra	Limited	Non-production only

Raise autonomy per area only after the system proves reliable there, by explicit founder decision recorded in DecisionQueue.md.