Sub Agent

Run saved assistant instructions in the background and inspect their results.

A scheduled Sub Agent runs a saved assistant instruction against your workspace in the background. It appears as Subagent on the Automations page. Each occurrence can make multiple model and tool calls and produces one run record with its final output and available tool details.

Chat is the interactive assistant surface. Automatic delegation from an interactive Chat response to a separate Sub Agent is currently disabled. Create a scheduled prompt task for background assistant work.

Creating a Sub Agent

Ask Chat to schedule an instruction, or select New Task in Automations. This modal creates Subagents only; workflow schedules, polling triggers, and webhooks are configured in their workflows. A task can run at a set time, repeat on a cron schedule, or repeat until a completion condition is met. You can choose the same models and supported reasoning levels as Chat. See Automation & Configuration.

External MCP clients can create these tasks with steelengine_jobs.create. The public MCP catalog exposes scheduled jobs through that domain tool; it does not expose an immediate natural-language delegation tool.

How a Scheduled Run Works

Temporal starts one worker activity for the complete assistant run. The worker loads the saved task and actor, checks current workspace access, resolves the model and provider key, and runs the shared assistant loop. Platform and integration tools execute directly in the worker. The initial run does not go through the web assistant-execution endpoint.

If the saved model loses its provider key or is blocked by the deployment, the worker resolves it to the workspace's default Chat model. Reasoning-effort and access checks still apply. Configure provider access on the Worker as well as the app; see Scheduled Assistant Configuration.

The prompt contains your saved instruction, task identity, completion condition when configured, and the last ten job-history summaries. The assistant can use update_job_history to record meaningful work for future runs. An until-complete task can call complete_job when its success condition is met.

Each run also receives its resolved model name and provider, including when the saved choice falls back to the workspace default. If your instruction asks the Sub Agent to identify its model, it uses the current run's identity rather than model names in previous job summaries.

Open the task's run history for status, duration, final output, tool details, and errors. A workflow started by the assistant creates a separate workflow execution record. Every successful scheduled run records a concise assistant note, separate from its final output, and explicitly decides whether the saved task is complete. An until-complete task calls complete_job when its success condition is met; without a separate condition, completing the saved instruction is sufficient. Unfinished and persistent tasks continue on their schedule, subject to the configured run limit. A run that cannot save its history or required completion is reported as failed, with the available output and tool details retained.

Selecting Run now shows Starting… immediately, then Queued until execution begins. Once the worker starts, the task and its run history show Running, including after a page refresh. The task detail updates automatically until the run finishes. A task that has stopped future runs still shows its current execution as running until that execution ends.

Model cost is saved on each new run's log and trace using the reported token usage and SteelEngine's model pricing, including cached input rates. Run history shows the total in credits; log details show model input cost, output cost, and token counts. If a run fails or is cancelled, usage reported by completed model steps is still included. When per-call usage is available, long-context pricing is calculated separately for each model call, rather than applying a larger-input tier to the whole run's combined tokens. Subagent model costs do not include a workflow base execution charge. Older logs without saved cost do not have a cost breakdown.

The activity does not checkpoint after each tool call or automatically retry the whole run. If it is interrupted, some external actions may already have completed. Inspect the run and destination system before starting it again. Job history helps the assistant avoid repeated work; it does not guarantee that external writes happen only once.

Authorization and Actions

A scheduled Sub Agent acts as the user saved on the task. Workspace and organization permissions, model restrictions, credential access, and plan rules are checked when it runs. Resource references and previous runs do not bypass those checks.

Scheduled runs have no interactive Approve / Deny prompt. Integration writes can run with accessible credentials under the task's instructions. A scheduled run can run a deployed workflow. It runs a saved workflow draft only when no block in the draft could change anything, because nobody is there to approve a draft that writes, sends or deletes. It cannot resume a run that is paused at a Human in the Loop step, because that step waits for a person's decision. Platform operations that require confirmation still require the tool's explicit confirm argument; this records caller confirmation, not a human approval event. Database deletion is unavailable to Sub Agents, Chat, and MCP, even if the actor can delete an empty database from the dashboard.

What You See in Interactive Chat

Chat renders each action inline, in the order it happened:

  • Tool rows: a status icon, a short title such as "Creating workflow", "Querying table", or "Searching knowledge", and a detail fragment naming the target. The icon animates while the call runs. It then shows one of four states: success, a warning for a partial result (with the skipped items and input errors listed), an error, or a stop mark. The stop mark appears when you stopped the response or denied the call. Click a row to expand the parameters and the result as JSON.
  • Approval prompts: destructive or confirmation-required actions pause with "… needs your approval to run." and Approve / Deny buttons. After you answer, the row records "approved" or "denied". A denied call never runs.
  • Connect prompts: when an integration tool fails because a credential is missing or needs reconnecting, the row shows a connect action that opens the OAuth flow.
  • Thinking blocks: for complex requests, an expandable block shows the model's reasoning while it works.

These tool rows describe work within the interactive response. A scheduled Sub Agent has its own task and run history.

Context

Chat receives the relevant context for the surface where it is opened. For example:

SurfaceContext
New chatAttached files, mentioned docs, tables, workflows, and prior chats
Workflow chatThe current workflow, selected blocks, workflow logs, and attached resources
Logs assistantThe logs visible on screen plus the workspace and workflow identifiers needed to answer follow-up questions

With each message, Chat also receives the current date and a bounded inventory of the workspace: the names and IDs of connected accounts, tables, knowledge bases, workflows, skills, MCP servers, custom tools, scheduled tasks, apps, and docs. The inventory lists only what you can read, up to a fixed number of entries per kind. It is not a snapshot of every resource and does not include resource contents. Chat looks up anything else with tool calls, and each lookup appears as a tool row. A Chat surface that cannot use tools receives the date only, and a scheduled Sub Agent receives neither.

When Chat creates a scheduled task, it can save the attached resource references, workflow branch and block selection, and captured view. Each run refreshes accessible resources by ID. An attached past chat supplies a bounded transcript from that exact chat, even if another chat has the same title. Captured views remain historical; later source-conversation messages do not change the saved instruction. Missing, unauthorized, or truncated context is identified explicitly.

Resource Work

Tool rows cover every SteelEngine resource Chat can act on:

Workflows

Build, edit, run, debug, deploy, and organize workflows.

Docs and Files

Read attachments, create docs, edit files, and save generated outputs.

Tables and Databases

Create tables, inspect schemas, query rows, and import structured data.

Logs

Inspect workflow, job, and execution logs and explain failures.

Knowledge Bases

Create knowledge bases, add documents, query content, and connect them to workflows.

Configuration

Help with credentials, environment variables, MCP servers, scheduled tasks, and deployment settings.

Relationship to SteelEngine MCP

Scheduled Sub Agents and interactive Chat use the same platform gateway and integration executor. SteelEngine MCP adapts the platform gateway for external clients such as Claude, Cursor, and VS Code.

They use the same underlying tool catalog, but they are different surfaces:

  • Sub Agent: a scheduled assistant run, configured and inspected in Automations
  • SteelEngine MCP: what an external MCP client calls programmatically

Legacy Names

Use these names when writing new docs and UI:

Legacy nameUse instead
MothershipChat, or "workspace-wide Chat" when the workspace scope matters
Workspace AssistantChat
CopilotChat (in the workflow editor)
Mothership tool callChat tool call
Copilot MCPSteelEngine MCP

Common Questions

No. One scheduled occurrence runs the complete assistant loop and counts as one run, even when the assistant makes several model and tool calls.
Yes, as a scheduled prompt task. Ask Chat to schedule the work, or call steelengine_jobs.create from MCP. Immediate delegation from an interactive Chat response is currently disabled.
Open Automations, select the Subagent, and inspect its run history. It contains the final assistant output and available tool inputs, outputs, and errors. Job-history summaries are recorded separately for future runs.

On this page