Canonadr
ADR-2026-09-04-10 — Capability, Governance and Hermes remain separated
Status: ACCEPTED — 2026-09-04 (Founder Canon §18–20; Founder decision "no monolithic AI Orchestrator")
Context
Convenient designs collapse "decide what to do," "decide what may be done," and "get it done"
into one orchestrator. That produces an autonomous actor with unreviewable authority — the
exact failure the product must never have.
Decision
The responsibility split is frozen:
Twin / Workbench = what the user is trying to achieve
Capability Discovery = what executable options exist
Capability Fabric = how it could be done (provider-neutral resolution)
Governance = whether it may be done and under whose authority
Hermes = how approved execution is coordinated reliably
Reconciliation = whether reality actually changed as intended
- Governance outcomes: ALLOW · DENY · REQUIRE APPROVAL · DELEGATE / ALLOW WITH BOUNDED
AUTHORITY. Confidence never grants authority; the Twin cannot self-approve consequential
action. - Delegation is explicit, capability-scoped policy (BOUNDED_DELEGATION) — never implied by
capability discovery. - Hermes starts only after a plan is approved or policy-allowed; it coordinates dispatch, waits,
retries, timeouts, recovery/compensation, evidence capture, verification and reconciliation —
and never owns canonical truth. - Execution providers may include APIs, MCP tools, agents, browser runtimes, workflows, local
execution, AI models or human tasks. No product surface, agent, connector or capability
implementation bypasses the owner of canonical composition, authority or mutation.
Alternatives considered
- A monolithic AI orchestrator combining cognition + authority + execution — rejected:
unreviewable authority, unbounded blast radius, vendor lock-in of the operating model. - Governance as a workflow step after execution — rejected: it is a cross-cutting boundary
evaluated before any consequential action.
Consequences
Every capability declares version, contracts, authority scope, risk, execution/sync modes,
idempotency, evidence requirements, provider routes, health, timeout, retry, compensation and
verification behavior before it is executable. Middleware and gateways must not create implicit
approvals.