Skip to main content

Built-in Harness

Every Autoheal agent — pre-built or custom — runs on the same built-in harness. This page describes what that looks like from the outside: what appears on screen while an agent works, and where you can intervene.

The agent loops between reasoning and tool calls, delegating focused work to subagents. When a governance policy gates a call, the loop pauses until you decide.

While the agent works​

A run streams as it happens rather than appearing all at once.

  • Thinking / Working — a live indicator tells you which. Thinking means the agent is reasoning. Working means it is calling tools or writing its reply. Beside it, a counter shows elapsed agent runtime.
  • Reasoning — the agent's thinking streams into a block you can read as it forms. When the step finishes, it collapses to a summary line such as Thought for 12s, which you can reopen.
  • The reply — written token by token, so you can start reading before the agent finishes.
  • Progress — investigations also keep a running to-do list, so you can see what the agent has planned and what it has finished.

The agent manages its own working memory. On a long session, it automatically compacts older context to stay within the model's limits. Your transcript is never shortened. The full record stays exactly as it happened.

Tool calls​

Each time the agent uses one of your integrations, a tool call appears in the transcript, tagged with the integration's icon and a status:

StatusMeaning
RunningThe call is in flight
CompletedThe tool ran and returned a result
Returned errorThe tool ran, but the target rejected the request
ErrorA platform fault. The call may not have run
BlockedStopped by security policy
StoppedYou interrupted the run

Open a call to see its Parameters and Result in full, with long output expandable or viewable fullscreen. On a finished turn, tool calls move under a collapsed Used tools header. The reply reads as prose first, with the evidence one click away.

Subagents​

For focused work, such as investigating one hypothesis or verifying a finding, an agent delegates to a subagent. The subagent appears as its own named card rather than a plain tool call.

A subagent card shows what it was asked to do, its status, and how long it took once finished. While it runs, the card is expanded so you can watch its output stream. You can open any subagent to read its full transcript, the same way you read the main run.

When the agent pauses​

An agent stops and asks you in two situations. Both appear in a dock above the message box. The box tells you why it is blocked: Resolve 2 pending requests to continue…. If several are waiting, you review them one at a time. You can minimize a pause, but you cannot dismiss it.

Awaiting approval — a governance policy gated a tool call. You see what the agent wants to run, its arguments, and which policy held it, with a link to that policy. You can:

  • Allow once — this call proceeds
  • Allow for session — this and later matching calls proceed without asking again
  • Reject — the call does not run. You can add a reason, which the agent receives

Destructive tools ask a second time before running, because the action takes effect immediately and cannot be undone from that screen.

Needs input — the agent has a question. You get suggested answers as one-click options and a free-text box for anything else. A Skip option lets the agent infer the answer and continue.

There is no separate resume step: answering or approving is the resume. Afterwards, the transcript keeps the decision (Allowed, Allowed for session, Rejected, Answered, or Skipped). So anyone who reads it later sees what was decided.

Staying in control​

  • Interrupt the agent — stop a run at any point from the message box.
  • Send a message mid-run — type while the agent is working and your message is held, then delivered when the agent stops. It survives a page refresh.
  • Quote and ask — select any text in the transcript to quote it into your next message.
  • Attach files — add screenshots, logs, or other artifacts to a message.

Following a run​

An investigation carries a status that you can read at a glance. Statuses include Investigating, Needs Input, Paused · action needed, Root Cause Ready, Stopped, Resolved, and others. The three that need you escalate to an amber pill so they stand out in a list.

Run Details records who started the run, when, how long the agent ran, and time to first hypothesis. A separate insights view shows model calls and token usage per model.

Every tool call and approval decision is also recorded in the Audit Trail, searchable and exportable long after the run ends.