Provider best practices

Prepare your agent provider for paid traffic, from token reporting to recovery and shutdown.

Provider production checklist

Build and operate / Evidence you need

Engine adapter and token counting: Real prepared bounds, cumulative usage, cache accounting and definitive outcomes agree with the engine.
Dispatch, renewal and cancellation: No dispatch before a grant; denied renewal stops within the covered bound; cancellation reaches the actual job.
TLS and HTTP/2 ingress: Declared hostname/certificate, both native RPCs and all discovery routes work through the deployed proxy.
Resource limits: Decompressed messages and queues are bounded; slow consumers and saturation do not silently drop accounting evidence.
Authoritative session ownership: Concurrent requests and workers cannot reserve the same headroom or restore an old Control generation.
Durable history and recovery: Complete checkpoints, signed run evidence, accepted catalogues and unresolved operations survive restart.
Wallet and recipient custody: Login renewal is owned; operator journal is retained; recipient signing stays outside serving.
Chain observation and collection: Fresh canonical reads, synchronized clocks, pause checks and a margin sufficient for final reports and transaction execution.
Shutdown and failure handling: Draining stops grants; actual jobs terminate; unresolved runs/operations remain recorded.
Updates and observability: Applied configuration matches publication; traces correlate protocol IDs; hooks are not counted as payments.

Build and operate	Evidence you need
Engine adapter and token counting	Real prepared bounds, cumulative usage, cache accounting and definitive outcomes agree with the engine.
Dispatch, renewal and cancellation	No dispatch before a grant; denied renewal stops within the covered bound; cancellation reaches the actual job.
TLS and HTTP/2 ingress	Declared hostname/certificate, both native RPCs and all discovery routes work through the deployed proxy.
Resource limits	Decompressed messages and queues are bounded; slow consumers and saturation do not silently drop accounting evidence.
Authoritative session ownership	Concurrent requests and workers cannot reserve the same headroom or restore an old Control generation.
Durable history and recovery	Complete checkpoints, signed run evidence, accepted catalogues and unresolved operations survive restart.
Wallet and recipient custody	Login renewal is owned; operator journal is retained; recipient signing stays outside serving.
Chain observation and collection	Fresh canonical reads, synchronized clocks, pause checks and a margin sufficient for final reports and transaction execution.
Shutdown and failure handling	Draining stops grants; actual jobs terminate; unresolved runs/operations remain recorded.
Updates and observability	Applied configuration matches publication; traces correlate protocol IDs; hooks are not counted as payments.

Verify the public endpoint, then validate your engine integration and operating behavior in your own environment. Include disconnects, denied coverage, stale chain data, lost wallet responses and restarts. Unpaid CLI verification alone does not establish a funded exchange or production readiness.

Plan for the whole session

Keep readiness checks, collection, recovery and shutdown in your operating workflow. Validate your engine under disconnects, denied coverage, stale chain data and process restarts.