Remote Agent

Remote Agent Delegation

If the plugin is not installed or configured yet, use the a2a-delegation-setup skill first.

Use the remote_agent tool to delegate work to external A2A-compatible agents and manage delegated tasks. send is the only action to start a new remote turn or continue an existing one.

When to delegate

Delegate when:

• The task requires capabilities you do not have (e.g. code execution, web search, domain-specific APIs)
• The user explicitly asks to send work to an external agent
• A configured target agent is better suited for the request

Do not delegate when you can handle the request directly.

Tool

The remote_agent tool exposes five actions: list_targets, send, watch, status, cancel.

Choosing a target

Call list_targets first to discover available agents. Prefer target_alias over target_url — aliases are stable names configured by the user. If a default target is configured, you can omit target_alias from send.

Actions

list_targets

Discover configured targets.

{ "action": "list_targets" }

send

Send a request to a remote agent. Use it for:

• a brand-new remote task
• a follow-up turn on an existing remote task
• a related new task that references prior work
• a new task inside an existing remote conversation

{
  "action": "send",
  "target_alias": "my-agent",
  "parts": [
    {
      "kind": "text",
      "text": "Summarize the latest quarterly report."
    }
  ],
  "follow_updates": true
}

• parts (required): non-empty array of text, file, or data parts.
• target_alias: configured target name. Omit when a default target exists.
• continuation: canonical persisted follow-up contract from summary.continuation. Round-trip this subtree verbatim for send, watch, status, and cancel.
• task_handle: manual compatibility input for follow-up actions when you are not replaying a persisted continuation.
• task_id: manual compatibility continuation id for send; for follow-up actions it identifies the remote task only when no task_handle is available inside a persisted continuation. task_id continues an existing task.
• context_id: manual compatibility conversation continuation id for send. Use it with task_id, or by itself to start a new task in the same conversation, only when you are not replaying a persisted continuation.
• reference_task_ids: optional related task ids for send. reference_task_ids references prior tasks without continuing them.
• task_requirement: optional durability contract. task_requirement="required" forces explicit task creation or fails fast.
• follow_updates: when true, streams updates and returns the full event log. follow_updates=true means “stream the initial send”; it does not guarantee task creation unless task_requirement="required".
• blocking: only for non-stream send; do not combine it with follow_updates=true.

Preferred continuation forms:

{ "action": "send", "continuation": { "target": { "target_url": "https://my-agent.example/", "card_path": "/.well-known/agent-card.json", "preferred_transports": ["JSONRPC", "HTTP+JSON"], "target_alias": "my-agent" }, "task": { "task_handle": "rah_abc123", "task_id": "task-123" } }, "parts": [{ "kind": "text", "text": "Approved. Continue." }] }

{ "action": "send", "continuation": { "target": { "target_url": "https://my-agent.example/", "card_path": "/.well-known/agent-card.json", "preferred_transports": ["JSONRPC", "HTTP+JSON"], "target_alias": "my-agent" }, "task": { "task_id": "task-123" } }, "parts": [{ "kind": "text", "text": "Continue the task." }] }

{ "action": "send", "continuation": { "target": { "target_url": "https://my-agent.example/", "card_path": "/.well-known/agent-card.json", "preferred_transports": ["JSONRPC", "HTTP+JSON"], "target_alias": "my-agent" }, "conversation": { "context_id": "ctx-123", "can_send": true } }, "parts": [{ "kind": "text", "text": "Start a new task in the same conversation." }] }

Manual compatibility only:

{ "action": "send", "target_alias": "my-agent", "context_id": "ctx-123", "reference_task_ids": ["task-1", "task-2"], "parts": [{ "kind": "text", "text": "Start related work without continuing those tasks." }] }

Continuation safety

Interpret follow-up capability from result.summary.continuation, not from prompt text, message text, or other inferred context.

• response_kind is descriptive only. Keep follow-up logic anchored to summary.continuation.
• summary.continuation.target: canonical persisted routing contract. Persist summary.continuation verbatim and pass it back directly for machine follow-up.
• summary.continuation.task: trackable task continuity. Use it for follow-up send, watch, status, and cancel.
• summary.continuation.conversation: conversation continuity only. Use it only with send to start a new task in the same conversation.
• summary.target_* is descriptive only and no longer part of the machine follow-up recipe.
• Top-level compatibility aliases stay descriptive only. Do not infer task continuity from flat task_id, flat context_id, or other top-level summary fields.
• Branch on summary.continuation.task vs summary.continuation.conversation before choosing the next action.
• Never infer or synthesize summary.continuation.task from summary.continuation.conversation, session ids, run ids, prior prompts, or summary text.
• Do not call watch, status, or cancel from a result that has only summary.continuation.conversation.
• Do not poll from conversation continuity.
• If lifecycle tracking is required, fail fast when the peer returns only summary.continuation.conversation.
• Do not route summary.continuation back through channel a2a; inbound A2A channel delivery is separate from remote_agent continuation.
• Inbound A2A defaults to originRoutingPolicy: "suppress-generic-followup" to reduce accidental host replay through channel a2a; treat legacy-origin-routing as an escape hatch, not as a supported queued follow-up path.
• If you see A2A_OUTBOUND_DELIVERY_UNSUPPORTED, the host selected the wrong boundary. Return to persisted summary.continuation plus remote_agent.

const task = result.summary.continuation?.task
const conversation = result.summary.continuation?.conversation

if (task) {
  // Trackable task lifecycle.
} else if (conversation) {
  // Send-only conversation continuity.
}

Invalid follow-up example:

{ "action": "status", "context_id": "ctx-123" }

That is invalid because status, watch, and cancel require task continuity, not just conversation continuity.

status

Poll the current state of a delegated task.

{ "action": "status", "continuation": { "target": { "target_url": "https://my-agent.example/", "card_path": "/.well-known/agent-card.json", "preferred_transports": ["JSONRPC", "HTTP+JSON"], "target_alias": "my-agent" }, "task": { "task_handle": "rah_abc123", "task_id": "task-123" } } }

watch

Subscribe to live updates from a running task.

{ "action": "watch", "continuation": { "target": { "target_url": "https://my-agent.example/", "card_path": "/.well-known/agent-card.json", "preferred_transports": ["JSONRPC", "HTTP+JSON"], "target_alias": "my-agent" }, "task": { "task_handle": "rah_abc123", "task_id": "task-123" } } }

cancel

Cancel a running task.

{ "action": "cancel", "continuation": { "target": { "target_url": "https://my-agent.example/", "card_path": "/.well-known/agent-card.json", "preferred_transports": ["JSONRPC", "HTTP+JSON"], "target_alias": "my-agent" }, "task": { "task_handle": "rah_abc123", "task_id": "task-123" } } }

Task handles

After a successful send, the result usually includes summary.continuation.task.task_handle (prefixed rah_) when the remote peer exposes task continuity. task_handle is returned only when the peer actually created a task. Persist summary.continuation verbatim and pass it back directly for follow-up send/watch/status/cancel actions. Handles are process-local and expire after restart or TTL, but nested continuation still round-trips safely because it also carries durable summary.continuation.target plus summary.continuation.task.task_id. Treat flat send.task_id, send.context_id, and target_alias as manual compatibility inputs, not as a replacement for nested continuation round-tripping. If the result includes only summary.continuation.conversation, there is no task lifecycle to poll, watch, or cancel.

watch vs status

Use watch when the remote agent supports streaming and you want live incremental updates. Use status to poll a snapshot of the current task state. If you are unsure whether the target supports streaming, start with status.

Errors

• UNKNOWN_TASK_HANDLE — the handle is expired or invalid. Retry with the same nested continuation, or re-send. Manual callers may fall back to flat task_id plus target routing only when they do not have a persisted continuation.
• TARGET_RESOLUTION_ERROR — the alias or URL did not resolve. Call list_targets to check available targets.
• VALIDATION_ERROR — invalid parameters. Check required fields for the action.