Agent permissions
Deny-by-default capabilities and human approval tiers 0–3 — how autonomous agents are actually governed.
Last updated — 2026-08
Deny by default
Agents hold no standing permissions. Every capability is granted explicitly per role and per task; anything not granted is refused. Visitor input is treated as untrusted data, never as instructions.
What agents may do
Analyze, classify, route, estimate, and draft internal plans — and build inside per-task sandbox workspaces. No agent can send a client-facing message, set a price, or deploy anything on its own.
Tier 0 — observe & compute
Read-only analysis: intake classification, team routing, and estimation. Runs fully autonomously, and every step is written to the audit log.
Tier 1 — draft
Agents produce internal artifacts — build specifications, delivery plans, pricing models. Drafts never leave the system without passing the tiers above.
Tier 2 — human approval gate
Anything client-facing stops at the deal desk: pricing, proposals, and commitments wait in an awaiting-approval state until a human operator approves or rejects. Nothing is sent without a human yes.
Tier 3 — verified delivery
Approved builds execute in isolated per-task workspaces and must pass a two-layer verification gate — automated checks plus an independent reviewer pass — before a human sees the result.
Everything on the record
Every agent event is timestamped in the audit log. The autonomy metrics published on /autonomy are computed from those same records — not from marketing estimates.