AWS Policies for the Nightly Live-Mode Workflow

Two JSON documents, both with placeholders to replace (<ACCOUNT_ID>, <AWS_REGION>, <S3_BUCKET>, <S3_PREFIX>; the repository genial-labs-ai/system_agent_harness_aws is already filled in):

file	attach to	purpose
oidc_trust_policy.json	the IAM role’s trust policy	lets GitHub Actions on main of this repository assume the role via OIDC; nothing else can (aud and sub conditions).
iam_policy.json	the same role’s permissions policy	least privilege: invoke only the two configured models (via their cross-region inference profiles and the underlying foundation models in the profile’s destination regions, which Bedrock requires), describe models/profiles, and read/write only under one S3 prefix.

Notes:

• bedrock:InvokeModel* covers the Converse API; there is no separate bedrock:Converse action. The condition bedrock:InferenceProfileArn restricts the foundation-model statement so the models can only be used through a profile in your account.
• If you change AGENT_MODEL_ID / JUDGE_MODEL_ID, update both ARN lists. The destination regions of a us. profile are listed with aws bedrock get-inference-profile.
• Anthropic models additionally require the one-time use-case form (aws bedrock put-use-case-for-model-access) and AWS Marketplace subscription permissions for the first invocation in an account; see the README.
• The optional Bedrock Evaluations lab (Day 4) needs a separate service role for the evaluation job (bedrock.amazonaws.com trust, S3 read on the dataset prefix and write on the output prefix) plus bedrock:CreateEvaluationJob / GetEvaluationJob / ListEvaluationJobs and iam:PassRole on that service role for the caller. It is deliberately not in iam_policy.json: the workshop never submits an evaluation job unless STOCKROOM_CONFIRM_AWS_SPEND=1 is set explicitly.
• CloudWatch GenAI observability (Day 2, live) needs the X-Ray / CloudWatch Logs permissions listed in the AWS “Add observability to your Amazon Bedrock AgentCore resources” page; they are also outside this policy because they are not part of the CI gate.
• The trust policy’s sub is pinned to main. For PR-triggered live runs you would add repo:genial-labs-ai/system_agent_harness_aws:pull_request; the workshop intentionally keeps live runs off PRs.