Managed Agent Harness (Bedrock AgentCore)
Created: 2026-06-02 09:14
#quicknote
A managed agent harness productises Harness Engineering by replacing the bespoke agent build process with a configuration-based deployment. AWS launched such a harness for Bedrock AgentCore, documented through Heeki Park's work on his Loom agent platform. The managed approach trades flexibility for speed: a managed-harness deployment completes in roughly 20 seconds against about a minute for a hand-built agent.
Deploy-Time vs Run-Time Configuration
The harness exposes configuration at two points, where the deploy-time settings act as defaults and run-time settings override them.
- Deploy time (
create-harness): system prompt, model,max-iterations,max-tokens, tools (including remote MCP servers), and memory resources. - Run time (
invoke_harness): per-invocation overrides for the model and for attached tools, allowing an end user to select a preferred model from an allow-list or enable a connector on demand.
This mirrors the hand-built pattern in Loom, where an AGENT_CONFIG_JSON payload conditionally enables MCP servers, agent-to-agent (A2A) connections, and memory hooks at deploy time, while the invoke request carries model_id, connector_ids, and credentials for run-time overrides.
Credential Delegation Concern
Park surfaces an identity concern relevant to AI Agent Security. Capturing user credentials at the moment a user toggles on a connector imposes the least consent fatigue, because intentionality is highest at that point, and it avoids the failure mode of a long-running task pausing to request permission after the user has stepped away. He notes this single-hop credential capture is a simplification rather than a true solution to delegated authority; RFC 8693 token exchange (carrying an actor claim in the JWT) is the more complete mechanism. At the time of writing, AgentCore credential providers did not yet support per-user API keys, so per-user keys had to be injected via request headers instead.
References
Tags
#harness_engineering #agentic_ai #ai_agents #aws #bedrock_agentcore #mcp #ai_agent_security