Search Agent Fleet - Multi-Agent AI Search Orchestrator

This product requires an ongoing internet connection. The following ongoing external services are required, using API keys you supply:

• Google Gemini (API host generativelanguage[.]googleapis[.]com, or Vertex AI): used on every request, for the orchestrator's and specialists' reasoning.
• Tavily and Brave Search: web search queries.
• Semantic Scholar, arXiv, DBLP: research queries.

HuggingFace is not required at runtime. The embedding model is baked into the image and the container runs fully offline for embeddings. Your prompts are sent to Google to be processed.

Required IAM permissions on the AgentCore Runtime execution role. This product calls the AWS Marketplace Metering Service directly from inside the container to bill your subscription. The "Create default role" option in the AgentCore Runtime console does NOT include this permission. Attach this AWS managed policy to the execution role before launching:

• AWSMarketplaceMeteringFullAccess: required for billing (MeterUsage). Do not substitute AWSMarketplaceMeteringRegisterUsage, despite the similar name; it grants a different, unrelated action.

Without it, the container self-terminates at startup rather than running unmetered; this is expected anti-fraud behavior, not a bug.

A second, separate IAM permission is also required, not covered by the policy above. This product also verifies your subscription's agreement status directly from inside the container at startup, calling aws-marketplace:SearchAgreements and aws-marketplace:DescribeAgreement. No AWS managed policy bundles just these actions; attach this inline policy to the execution role:

{ "Effect": "Allow", "Action": [ "aws-marketplace:SearchAgreements", "aws-marketplace:DescribeAgreement", "aws-marketplace:GetAgreementEntitlements" ], "Resource": "*", "Condition": { "ForAllValues:StringEquals": { "aws-marketplace:AgreementType": ["PurchaseAgreement"] }, "StringEquals": { "aws-marketplace:PartyType": "Acceptor" } } }

Using Database Search. There is no connection-string configuration field: you include your database's connection string directly in your prompt text, e.g. "Connect to postgresql://user:pass@host:5432/dbname and list the tables". Supported: SQLite, PostgreSQL, MySQL, and MongoDB. Only read-only statements are permitted (SELECT / WITH...SELECT / SHOW / DESCRIBE / EXPLAIN for SQL); mutations are rejected. Your database must be reachable from the runtime's network; if it's private, run the AgentCore Runtime in VPC mode with connectivity to it. Connection strings are masked in logs and never echoed back in a response, but treat each prompt as sensitive since it carries live database credentials.

Using Vector Search durably. Without extra configuration, indexed documents are lost on every restart; fine within one session, not for a persistent knowledge base. To persist across restarts, mount an EFS access point into the runtime and point CHROMA_PERSIST_DIR at the mount path. Mounting EFS (or S3 Files) is required either way, since this product has no upload endpoint of its own.