APG documentation.
Everything to stand up the Agents Proving Ground runtime, register agents, and choose an inference tier. The quickstart is ready. We are still finishing the deeper guides, and until they are done the product repository is the source of truth.
The app running
Until they are done, the APG repository is the source of truth.
What the guide covers.
Quickstart
Download and unpack the release, run docker compose up -d, and run your first agent on your own key (Tier 2) or your own GPU (Tier 3). The SF-default tier is coming soon.
Architecture
Registry, runtime loop, validation boundaries, and the read-only control plane.
Self-host
Requirements, Docker and compose, and the data plane on your infrastructure.
Configuration
Cost ceilings, health checks, operator-confirm gates, and per-agent settings.
Agents
Registering agents, schemas, and the read-reason-propose-gate lifecycle.
API
Running agents, reading runs, and the control-plane endpoints.
Integrations
Binding CRM, Drengur, and third-party self-hosted apps to the runtime.
Operations
Inference tiers (SF default, BYOK, sovereign), health, and cost monitoring.
Troubleshooting
Failed validations, ceiling trips, and model-day degradation.
ADRs
The decisions behind the runtime, including the approval gate on outbound actions.
Tell us what the work is.
A call is the fastest way to find out whether an agent should carry a piece of your operation, and what it would take to run it on your own machines.