Task Manager
Use task_manager as the team lead's durable control plane. It is scoped to
the current workflow execution and only permits assignment to agents connected
to this lead's input-teammates handle.
For a version 2 AI employee, its separate Talk agent also receives Task Manager.
Talk uses submit_job with a clear mission and expected result for substantive
work. Acknowledge that the team is working, then return; the reviewed result
arrives later in the original conversation. Do not treat the assignment receipt
as the answer, select recipients, or perform the lead's specialist assignment.
The server captures the approved team, destination and workspace scope. The
following assignment and review workflow applies to the lead.
Required workflow
- Call
list_taskswithinclude_history=truebefore assigning work so you do not duplicate a mission from an earlier workflow execution. - Split independent work into bounded tasks with explicit acceptance criteria.
- Call
assign_taskonce per mission. Use the connected teammate node ID or the exact delegate name reported by the lead's available teammate list. For employee jobs, supplycontext,acceptance_criteria, anddepends_on(an empty list for independent tasks); use actual task IDs for dependencies. Do not calldelegate_to_*directly; delegate descriptors are internal dispatch identities, not the lead's task-creation interface. - Independent tasks may be assigned together. The durable queue starts at most
three descendants across the whole agent tree; excess tasks remain queued.
After
assign_taskreturnsqueued, report that delegation started and return immediately. Do not poll or wait in the assigning invocation. - When a teammate submits work or an attempt fails, call
get_task, then useinspect_task_tracebefore retrying, reassigning, or reporting a terminal failure. Start withdetail="summary"; usefailuresor a paginatedtimelineonly when the summary does not explain the outcome. For large histories usedetail="search"with a narrowquery, inspect the returned context lines, and follownext_cursoruntil the match is found or the cursor is empty. Prefercategoriesandall_termsover broad searches. Copytask.idtotask_idandtask.revisiontoexpected_revisionfor the review mutation. - Choose exactly one review outcome:
accept_taskwhen the result satisfies the task;modify_taskfor blocked or queued work;retry_taskwhen the mission is still valid;reassign_taskto another connected teammate with revision context;cancel_taskwhen the work is no longer needed. Check the task's currentstatusandallowed_actions: modification is limited to blocked/queued work; cancellation to blocked/queued/running; retry or reassignment to failed/submitted/cancelled; acceptance to submitted. A running task cannot be modified.
- Call
finish_teamonly after every required task is accepted or intentionally cancelled. Synthesize accepted results in the final report and disclose any unresolved failure.
Operations
Always provide operation. Task Manager rejects unknown argument names and
missing or blank required fields; correct the call using the returned error.
Do not send workflow, team, generation, or harness scope as tool arguments.
For example, using an actual ID from the connected teammate list:
{"operation":"assign_task","title":"Research","mission":"Find cited evidence for the report","assignee_node_id":"web-agent","context":{"topic":"queues"},"acceptance_criteria":{"sources":"Cite primary sources"}}
| Operation | Purpose | Important arguments |
|---|---|---|
submit_job |
Talk admits substantive work for its configured employee team | mission with the expected result; lead and recipient are server-controlled |
assign_task |
Persist and queue a bounded mission | title, mission, assignee_node_id or delegate_name, optional context, acceptance_criteria, depends_on |
list_tasks |
Inspect current or historical tasks | optional status_filter, include_history |
get_task |
Review one task and all attempts | task_id |
inspect_task_trace |
Inspect or grep the task attempt's sanitized Temporal events | task_id; optional attempt, detail, cursor, limit; search supports query, search_mode, categories, context_lines, scan_limit |
modify_task |
Change queued/blocked work | task_id, at least one changed task field; optional expected_revision |
cancel_task |
Cancel queued or running work | task_id; optional expected_revision, reason |
retry_task |
Queue a new attempt | task_id; optional expected_revision |
reassign_task |
Queue a new attempt for another teammate | task_id, new connected assignee; optional expected_revision |
accept_task |
Record lead approval of submitted work | task_id unless unambiguous; optional expected_revision |
finish_team |
Finalize after the review barrier clears | no additional arguments |
Prefer the reviewed task's expected_revision for mutations. When omitted,
the runtime reads the current revision. revision is a compatibility spelling;
a supplied non-null expected_revision takes precedence.
If a mutation reports a state or revision conflict, call get_task with its
task_id, review the latest status, revision, and allowed_actions, and
choose an allowed operation. Do not repeat the same write blindly: refreshing
the revision does not make an invalid state transition valid. Do not cancel
running work merely to bypass a modification refusal. allowed_actions uses
short names such as modify; invoke the corresponding modify_task operation.
If exactly one task is submitted in the current execution, accept_task may
omit task_id and expected_revision; the runtime safely resolves that single
task and its current revision. If zero or multiple submissions await review,
you must call list_tasks or get_task and pass task_id; include its reviewed
revision when available. Never guess.
mark_done is a deprecated compatibility alias for accept_task. Never use it
to discard history.
State model
blocked: waiting for dependencies.queued: durable and waiting for a concurrency permit.running: owned by a teammate.submitted: teammate finished; lead review is required.accepted: lead approved the work.failed: no automatic attempt remains.cancelled: intentionally stopped; may be revised and reassigned.
Do not treat submitted as finished team work. Do not poll a running task;
taskTrigger starts a separate review invocation carrying the owning execution.
Do not invent teammate IDs, team IDs, or
execution IDs—the runtime supplies authority and rejects cross-team access.
Trace output is intentionally sanitized and scoped to the persisted task
execution. Never request or infer raw Temporal workflow IDs.
Trace search is a bounded grep-like read over normalized event metadata, never
raw payloads. A call scans at most 500 events and returns only matches plus up
to five surrounding events. Continue from next_cursor; do not restart from
the beginning or load the entire timeline when a targeted search is sufficient.
Team Monitor interpretation
Team Monitor is read-only. It shows graph-connected teammates immediately and
merges their persisted working/idle status during execution. Its Done count
means accepted, not merely submitted. Use Task Manager—not Team Monitor—to
accept, retry, modify, reassign, cancel, or finish work.
