Trace AI subagents

Trace AI can delegate independent specialist work to configured Autonomic Agents. Delegated work runs asynchronously, so Trace AI can start several independent tasks, continue the conversation, inspect progress, and cancel work that is no longer needed.

What a subagent is

A subagent is one isolated run of an existing Autonomic Agent definition. It is not a separate chat agent and does not appear in the agent picker.

Each delegated run has:

• its own hidden child session and transcript;
• a run-specific workspace and backlog;
• a parent link to the Trace AI conversation;
• ordered progress events;
• an inspectable final decision;
• queued or cooperative running cancellation.

The catalog always includes repo-wired platform specialists (platform-bulk-reader, platform-code-specialist). You do not create those under Agents first. Tenant-authored Autonomic Agents appear only when a tenant administrator enables Allow Trace AI to delegate specialist runs.

The deployment must enable both AUTONOMIC_AGENTS_ENABLED=1 and TRACE_SUBAGENTS_ENABLED=1.

Delegate work from Trace AI

Ask Trace AI naturally, for example:

Split the ProcessFlow step implementation and WebApp review into independent specialist tasks. Run them in parallel and summarize the validated results.

Trace AI loads load_platform_skill with skill_id: subagents, checks list_subagent_catalog, and delegates only to matching specialists. Prefer a source: "platform" slug when it fits.

Good delegated tasks are:

• self-contained;
• independent of unresolved work;
• explicit about inputs and constraints;
• explicit about the definition of done;
• explicit about the expected result.

Trace AI remains responsible for clarification, overall planning, verification, and the final answer.

Progress and completion

Trace AI receives a run_id immediately. It can:

• inspect one run with get_subagent_run;
• list runs linked to the current conversation with list_subagent_runs;
• read ordered tool and lifecycle events;
• read interruption metadata when a run is blocked (orchestration vs human);
• reconcile runs in a later turn if they outlive the current response.

When interruption.kind is orchestration (need_parent), Trace must stop polling, supply missing delegation context, and re-delegate. When it is human (need_human), Trace mediates the question to the user instead of routing through the Autonomic Agents screen alone.

Delegated runs do not appear as separate conversations in normal chat history.

Cancellation

Ask Trace AI to cancel a superseded delegated task.

• A queued run is removed from the ProcessFlow queue immediately.
• A running run changes to cancel_requested and stops at the next safe model or tool boundary.

Running cancellation is cooperative. It cannot interrupt an external operation that has already started.

Platform specialists (no setup)

Official jobs shipped in the product:

The first delegate_subagent to a platform slug compiles the tenant Autonomic Agent. Full operator page: Platform specialists.

Code-generation specialists

The platform code-generation capability covers:

• ProcessFlow process-step snippets in TypeScript, JavaScript, or the existing PHP compatibility mode;
• WebApp HTML, CSS, JavaScript, and HTTP headers.

platform-code-specialist unlocks skill_id: platform-code-generation on the child turn (no processes / webapps mutation tools on the child). Trace applies validated artifacts with domain skills after Phase 1 completes. A tenant-authored specialist should load platform-code-generation explicitly and follow the same artifact-only / Trace-mutates split unless the mandate says otherwise.

Read-only validation does not save, publish, or execute the generated code. Those actions remain separate and keep their existing authorization and confirmation rules.

Repository checkout, arbitrary shell execution, package installation, and Git operations are not included.

Configure a tenant specialist

Tenant administrators add extra specialists under Agents (not required for the two platform slugs):

1. Create or edit an Autonomic Agent.
2. Set a focused mandate (Purpose, Context/fitness, Sense-Reason-Act-Evaluate, how fitness is ensured, stop rules) and activate the agent.
3. Enable Allow Trace AI to delegate specialist runs.
4. Add a short delegation description.
5. Set maximum concurrent delegated runs from 1 through 4.
6. Keep process and peer allowlists minimal.

Do not create Autonomic Agents whose slug is platform-bulk-reader or platform-code-specialist unless you intend to replace the official template for this tenant.

See also

• Trace AI Agent
• Autonomic Agents
• Platform specialists
• Platform code generation
• Platform skills quick reference