Skip to main content
The side panel shows what happens behind an answer: the plan and its progress, the subagents the agent dispatched and how they connect, the pages the agent’s browser loaded, every tool call with its input and output, and how much context and time the run used.

Open the side panel

In a conversation, click Open side panel at the top right of the chat. A blue dot on the button means a subagent is still queued or running.
  • Large screens (1024 px and wider): the panel opens beside the chat. Drag the divider between them to resize it, or focus the divider and use ← →. Double-click the divider to go back to the default width. The width you set, and whether the panel is open, are kept in this browser.
  • Smaller screens: the panel opens full screen over the chat.
The panel opens on the tab you used last, or, if that tab has nothing to show, on the first tab that has. Click Close side panel at its top right to close it.

Tabs

The tab bar shows each tab’s count while there is something to count: On narrower panels the tabs show only their icon and badge; hover over one to see its name.
The Plan tab shows the goal of the current task and the steps of its plan.
  • Goal: the outcome OASM pinned for the task. It reads Goal · done once every step is completed or failed.
  • Steps: each step with its status, a progress bar, and a count such as “4/4 done”.
Side panel Plan tab with a Steps section marked 4/4 done, a full green progress bar and four struck-through completed steps, next to the agent's final remediation answer

The Plan tab after every step is done

On medium-sized screens the plan also sits, folded, above the composer. Click it to unfold it; folded, it shows the step in progress.With no plan yet, the tab reads “No plan yet”. Ask mode never writes one. See Plans and approvals.

How subagents work

In Agent mode the main agent orchestrates: it reads workspace data, keeps the plan, and hands the hands-on work, such as scanning, crawling, testing one surface or verifying one finding, to subagents. Each subagent is one of the agents available to you, given one focused job.

Dispatch

For each job, the main agent writes a brief: the objective with its exact targets, what to hand back, what is already known, what to do and what not to do. A subagent sees only its brief, its role’s instructions and the newest notebook entries, so the brief has to stand on its own. Its targets must be inside the scope, or the dispatch is refused (see Safety checks). Each subagent has its own context window and step budget. A long scan costs the main agent one structured result instead of pages of raw output.

Parallel and chained work

  • Independent jobs are dispatched together and run side by side. A conversation can have up to 10 subagents queued or running at once. Each OASM API server runs up to four subagents at the same moment, across all conversations; the rest wait as Queued.
  • A job that needs another job’s output is dispatched right away with a Waits for link. It starts when those jobs end, whether they succeeded or not, and receives their results directly.
  • A subagent can get a follow-up message from the main agent (a Follow-up link), for example when its result was partial or raised a question.
  • A subagent that runs longer than 20 minutes is stopped and marked failed.

Results and waiting

When a subagent finishes, it hands back a structured result: an outcome (done, partial or blocked), a summary, candidate findings, facts and open questions. Facts and open questions go into the notebook. The result reaches the main agent on its own at its next step; the main agent does not poll. When nothing else is left to do, it waits for the running subagents, up to about eight minutes per message. Results that arrive later are handed to it at your next message. A plan step that dispatched subagents is complete only once their results are back and read, not when they were dispatched.

What subagents may run

Subagents run unattended: no approval prompt reaches you from them.
  • A built-in agent runs the commands and tools its role needs without asking, in both approval modes.
  • A custom agent gets read-only tools in a Manual chat, and the tools it declares in an Auto chat.
  • The safety checks still apply, and a call they would escalate to a prompt is refused.
  • Subagents cannot dispatch subagents of their own.
See Plans and approvals.

Stopping

  • Stop in the composer ends the run and cancels every queued and running subagent of the conversation. Their rows turn Cancelled.
  • Stop in an agent’s details on the Workflow tab cancels that one subagent only.
  • Closing the browser tab stops nothing: the run and its subagents keep going on the server.

Chat

The composer, modes and conversations

Plans and approvals

Approve plans and commands

Agents

The built-in and custom agents subagents are drawn from

Usage

Tokens and cost across all your conversations