Agent Worker Data Model

The portable agent workers use the public Conductor Task and TaskResult types. The task output is the durable checkpoint shared by embedded execution and remote SDK polling.

1. AgentClient

AgentClient is the control-plane interface used by the conductor branch. Its methods cover compile, deploy, start, status, execution search, respond, stop, signal, cancel, and SSE streaming. The worker depends on this interface rather than on an HTTP controller or core WorkflowExecutor.

2. ConductorAgentRequest

ConductorAgentRequest extends AgentStartRequest. It therefore supports the controller's start fields, including:

Field	Type	Purpose
name	String	Previously deployed agent name
version	Integer	Optional deployed version
prompt	String	User input for start or resume
model	String	Per-call model override
sessionId	String	Session association
media	List<String>	Media inputs
context	Map<String,Object>	Additional execution context
idempotencyKey	String	Optional caller-supplied key
framework / rawConfig	framework-specific values	Inline foreign-agent construction
runId	String	Per-execution worker-domain isolation

The worker adds:

Field	Type	Purpose
executionId	String	Resume an in-flight run instead of starting a new one
pollIntervalSeconds	Integer	Poll delay; default 5
maxDurationSeconds	Integer	Absolute deadline; default 86400
maxPollFailures	Integer	Consecutive transient-failure cap; default 30

3. ConductorAgentExecution

ConductorAgentExecution is the normalized snapshot consumed by ConductorAgentDelegate:

Field	Meaning
executionId	Agent execution identifier
agentName	Executed agent
state	Normalized ConductorAgentState
output	Structured result
text	Latest or final text
pendingTool	Pending human/tool request
reasonForIncompletion	Failure or cancellation reason
startTime	Agent execution start time in epoch milliseconds
endTime	Terminal agent execution end time in epoch milliseconds

States are RUNNING, WAITING, COMPLETED, FAILED, and CANCELED.

4. Durable task output

ConductorAgentResults owns the persisted keys:

{
  "executionId": "exec-xyz",
  "agentName": "research_agent",
  "state": "working",
  "taskId": "exec-xyz",
  "contextId": "session-123",
  "task": {
    "kind": "task",
    "id": "exec-xyz",
    "contextId": "session-123",
    "status": { "state": "working" },
    "metadata": {
      "agentType": "conductor",
      "executionId": "exec-xyz",
      "agentName": "research_agent"
    }
  },
  "agentStartTime": 1752451200000,
  "agentPollFailures": 0
}

The shared state, taskId, contextId, task, agentMessage, artifacts, and text fields use the same A2A v0.3 wire model as remote A2A agents. Native states map as follows: RUNNING to working, WAITING to input-required, COMPLETED to completed, FAILED to failed, and CANCELED to canceled.

Terminal executions additionally expose agentEndTime. Completed executions retain the legacy output map and add A2A message/artifact parts, including a data part with the full structured result. Waiting executions complete the current task with waiting: true, an A2A input-required status, and may expose pendingTool and text.

5. Invariants

1. agentStartTime is set once and survives retries and process restarts.
2. executionId is persisted before later invocations poll status.
3. A successful status call resets agentPollFailures to zero.
4. The idempotency key uses retry-stable task identity, never taskId.
5. Every invocation returns a complete TaskResult; the annotation runtime applies and persists it.