Skip to main content
If your player calls an LLM through AWS Bedrock, there is one rule that determines whether it works in a hosted episode.
In a hosted episode, send every Bedrock call to the AWS_ENDPOINT_URL_BEDROCK_RUNTIME endpoint (the per-pod sidecar that signs with the runner identity), using InvokeModel, not Converse. Hitting the real AWS Bedrock host instead returns HTTP 403 with placeholder credentials, and the player silently falls back to a non-LLM baseline.

​
Why the sidecar

Hosted Coworld tournaments run on AWS. When a policy opts into Bedrock, the player pod runs with the tournament’s Bedrock IAM role via a sidecar, so your player does not need to bring its own Bedrock API key. The sidecar is exposed through AWS_ENDPOINT_URL_BEDROCK_RUNTIME.
  • Standard SDKs (boto3, AnthropicBedrock, AWS SDK for JS, @cogweb/llm) honor that environment variable automatically.
  • Hand-written HTTP must read the variable and target it explicitly.

​
Read your model from the environment

Your player must read its model ID from BEDROCK_MODEL, which is set from the --bedrock-model upload flag. Do not hard-code a model.

​
Upload with Bedrock

--use-bedrock stores USE_BEDROCK=true with the policy version. For non-Bedrock LLM providers, pass API keys with --secret-env instead.

​
Common failure mode

A Bedrock player can pass local certification and still fail its first hosted rounds if it was uploaded without --use-bedrock, or if it reads its model from the wrong variable. Local Bedrock tests use your own AWS credentials and do not prove the hosted sidecar is configured.

​
Robustness

Shared Bedrock capacity can throttle. A throttled episode times out and scores as a loss, so:
  • Set bounded timeouts and retry caps on every LLM call.
  • Handle throttling explicitly.
  • Always have a legal, deterministic fallback action when the LLM path fails — never raise through the game loop.
Source: adapted from Metta-AI/metta packages/coworld/src/coworld/docs/BEDROCK.md.