@agent-dispatch/worker-agentcore

Reference AWS AgentCore worker runtime for AgentDispatch.

This package is the cloud-side process you can package into an ECR image and run through AWS AgentCore. It receives tasks from @agent-dispatch/adapter-aws-agentcore, executes them with your chosen agent framework, and exposes protocol endpoints that let the lead agent continue interaction after spawn.

What it provides

• HTTP task endpoint for agent.run and command.run payloads.
• A2A-compatible JSON-RPC message/send endpoint.
• Agent Card discovery at /.well-known/agent-card.json.
• Health endpoint for runtime checks.
• Pluggable execution boundary for OpenClaw, Hermes Agent, LangChain, Strands, custom scripts, or direct model calls.

Runtime contract

Endpoint | Purpose --- | --- GET /ping | AgentCore health check. Returns {"status":"Healthy"}. GET /health | Local liveness alias. GET /.well-known/agent-card.json | Returns A2A agent metadata and supported skills. POST / | Accepts A2A JSON-RPC message/send messages. POST /invocations | Accepts the AgentDispatch HTTP task envelope.

The default image is optimized for AgentCore A2A runtimes. It listens on port 9000 unless you override PORT, matching the AgentCore A2A service contract. For HTTP envelope mode, run with AGENTDISPATCH_WORKER_PROTOCOL=http PORT=8080.

The default worker is intentionally small. It proves the AgentCore runtime path and gives teams a place to wire their real subagent implementation.

Run locally

npm install
npm run build
AGENTDISPATCH_WORKER_PROTOCOL=a2a npm start

Send an A2A-style message:

curl -X POST http://localhost:9000/ \
  -H "content-type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "id": "1",
    "method": "message/send",
    "params": {
      "message": {
        "role": "user",
        "parts": [{ "kind": "text", "text": "Run the background task." }]
      }
    }
  }'

Build a runtime image

docker build -t agentdispatch-worker-agentcore .
docker tag agentdispatch-worker-agentcore:latest 123456789012.dkr.ecr.us-west-2.amazonaws.com/agentdispatch-worker:latest
docker push 123456789012.dkr.ecr.us-west-2.amazonaws.com/agentdispatch-worker:latest

Use the pushed image in @agent-dispatch/adapter-aws-agentcore runtime mode with ecrImageUri and executionRoleArn.

The AWS adapter automatically sets AGENTDISPATCH_WORKER_PROTOCOL on runtime-mode deployments to match the requested AgentCore protocol. You can override or add environment variables with target.details.environmentVariables.

Customizing the worker

The default executor is echo, which is only a smoke-test implementation. For a real cloud subagent, configure a command-backed framework adapter in the runtime environment:

AGENTDISPATCH_AGENT_FRAMEWORK=openclaw
AGENTDISPATCH_FRAMEWORK_COMMAND_OPENCLAW="openclaw run --stdin-json"

Or configure multiple frameworks at once:

AGENTDISPATCH_FRAMEWORK_COMMANDS='{
  "openclaw": "openclaw run --stdin-json",
  "hermes": {
    "command": "hermes-agent run --stdin-json",
    "timeoutSeconds": 1800,
    "env": {
      "HERMES_MODE": "subagent"
    }
  }
}'

When input.framework is openclaw, the worker launches the configured command without a shell, sends this JSON envelope to stdin, and maps the command response back into AgentDispatch events and results:

{
  "taskType": "agent.run",
  "framework": "openclaw",
  "instruction": "Run the background task",
  "context": {},
  "input": {},
  "metadata": {}
}

Framework commands can return either plain text or structured JSON:

{
  "output": "Task completed.",
  "result": {
    "summary": "Task completed."
  },
  "events": [
    {
      "type": "task.progress",
      "message": "Repository scan complete."
    }
  ],
  "artifacts": [
    {
      "uri": "s3://bucket/task/result.json",
      "kind": "json",
      "contentType": "application/json"
    }
  ]
}

This makes the package usable with:

• OpenClaw or Hermes Agent for native subagent execution.
• LangChain, Strands, or custom orchestrators for tool-heavy workflows.
• Direct model calls for simple long-running analysis.
• Internal tools exposed by the runtime container or cloud environment.

The important boundary is the protocol contract: accept an AgentDispatch task or A2A message, run the subagent, emit structured results, and keep cloud credentials inside the runtime environment.

Development

npm install
npm run typecheck
npm test
npm run build

See the release workflow for npm Trusted Publisher setup, provenance publishing, and upstream package order.