Agent Traffic Control (ATC)
Status: live. A public interactive demo runs at griff.run/atc, and a full 76-minute governed agent session is published with hash-backed replays at griff.run/audit.
Page updated 2026-08-10 against the live platform.
ATC routes every agent action through one governed path:
- Request — an agent asks to do work; the request is captured with an owner and boundary.
- Load context — GRIFF Brain retrieves the relevant memory, documents, decisions, and operating instructions.
- Evaluate — the policy engine determines what the agent is trying to do and which controls apply.
- Approve — sensitive classes (external sends, spend, credentials, production changes) pause for a named approver. Everyday work keeps moving.
- Execute — the agent performs the approved task with the capabilities available to it.
- Record — request, context, approvals, actions, outputs, and outcome are recorded as work happens.
- Review — managers, security teams, customers, and auditors inspect or replay the record when needed.
What ATC enforces out of the box
| Control | Behavior |
|---|---|
| Governed routing | No ungoverned back doors; one path for every request. |
| Human approval gates | Sensitive action classes pause for a named approver. |
| Shared operating context | Claude, ChatGPT, Gemini, Grok, custom agents, and MCP apps work from the same current state via GRIFF Brain. |
| Evidence by default | Replayable, exportable records for reviews and audits. |
| Budget & usage control | Model calls attributed per workspace; spend is a governed input. |
| Deployment freedom | The same model runs cloud-connected, hybrid, or self-hosted. |
Where you see it
- Control Center — approvals, sessions, and usage in the customer workspace (app.griff.run).
- Audit replay — the published evidence session at griff.run/audit.
- Governed local runtime — on-platform agents call the Brain's MCP surface where every governed call receives a policy decision and an audit entry (see the MCP API reference).