Watching & steering the fleet

Most of the time Harmonic runs without you. This page is about the times you look in, and the times it asks you to.

The board

Each Workspace has a board, one card per ticket, that mirrors your tracker. At a glance you can see what’s waiting, what’s running right now, what finished, and what got handed back to you. It’s built to sit on a side monitor: you watch the queue drain and only lean in when a card asks for it.

Watching a ticket work

Open any running ticket to watch its agent live. As it works you see what it’s reading, the changes it’s making, its plan, and its reasoning, streamed as it happens, with running token usage and cost. You don’t have to watch, but when you want to know why an agent did something, it’s all there rather than buried in a log after the fact.

There’s also an instance-wide activity view that shows every agent running across all your Workspaces at once, with the same live usage and cost, so you can see the whole fleet’s load in one place.

The timeline

The board shows where every ticket stands right now; the timeline shows what the fleet has been doing. It lays every attempt each harness has run onto one clock, so you can see what overlapped, what took a while, and when it happened. Scrub the playhead back to any moment to read the fleet’s state then, or open an attempt to step through its own run. Choose a 24-hour or 7-day window. Each unattended Attempt also records its effective permission mode here. If a requested mode was unavailable, the timeline shows the requested-to-effective fallback rather than hiding the change.

When a ticket needs you

A ticket comes back to you in one of two ways, and both are marked clearly on the board rather than failing silently:

• Escalated. The agent hit something only a human can settle, a permission it needs granted, or a question it can’t answer, so Harmonic stopped it and flagged the ticket for you. It won’t run automatically again until you take it on.
• Out of retries. The work didn’t hold up and Harmonic exhausted its automatic retries. The ticket is left open and flagged for you to look at.

Retry, Accept and Close

An escalated ticket has three actions on its page. They sit in a bar across the main column, directly under the Escalated banner. Hover over Accept to see a short explanation of what it will do at the current step.

• Accept overrides the step that failed and carries on with the pipeline. Its label names that step: Accept & implement after a failed rebase, Accept & verify after a failed implementation, Accept & review after a failed verification, and Accept & merge after a failed review. If no step is known, it reads Accept. Accepting at the final review merges the candidate as it is. Accept is disabled, with No candidate commits to accept, when there is nothing to accept.

• Retry sends the ticket back to the queue on the same branch, and the attempt budget starts over. The dialog has these fields:

  • Guidance (optional): what was wrong and what the next Attempt should do differently. It is recorded on the escalated Attempt and given to the next one.
  • Harness and Model: the route the next Attempt uses. Choosing a different one saves it on the ticket, which replaces the ticket’s Routing Label. The dialog says when a label will no longer apply.
  • When: Retry queues the next Attempt, and it starts when a slot is free. It continues the current Session only if that Session is still warm and the Harness and Model are unchanged. Retry Now starts the next Attempt at once, without waiting for a free slot. When there is a prior Session to re-use, Re-use the same Session appears; it is unavailable when the Harness changes, because a different Harness needs a new Session. The dialog warns when the Session is cold, or when the Model is switched within it, because the cached context can’t be reused then and the run costs more.

  The confirm button is labelled to match: Retry, Retry Now, or Retry Now in this Session when the Session is re-used.

• Close ends the ticket without merging its candidate. It removes the branch and worktree and closes the tracker issue. It asks you to confirm first, and it cannot be undone.

An escalated Epic has its own panel. See Epics.

Taking over

When you pick up a flagged ticket, you’re working it by hand, with the same agents, through Matt Pocock’s Skills directly, or however you’d normally resolve it. Answer what the agent couldn’t, and either finish it yourself or hand it back for another automatic run once it’s unblocked.

Steering, pausing, and resuming a ticket

You don’t have to wait for a ticket to come back to you. Type into the steer box on a running ticket and your message goes straight to the agent’s current conversation, or waits for the end of its current turn if it can’t be interrupted.

Pause asks the agent to finish what it’s doing and stop, keeping its work and its conversation. Resume picks the same attempt back up. Steering a paused ticket resumes it and delivers your message in one step.

A pause can last as long as you like. If the saved conversation can’t be reloaded as it was, for example after you upgrade Harmonic, the agent starts again from a summary of the work so far. Resuming a ticket that’s been left a while costs more, because the provider’s cache has gone cold, but it’s never refused. The same goes for a ticket whose agent process has gone: steer or resume it and it picks up again.

If a running ticket is close to its time limit, Extend gives it more time without restarting it. The ticket shows how much time it has left.

Pausing everything

The Pause button in the header freezes every running ticket in every Workspace, and anything that starts while it’s on is paused straight away. Press Resume and paused tickets carry on. The Auto-runner switch next to it is separate: turning it off stops Harmonic picking up new work, but leaves running tickets alone. Per-Workspace throughput dials (priority and concurrency) live in Settings & overrides.