Watch Your AI Agents Work — and Step In Whenever You Want

Release v1.0.165

Autonomy is only valuable when it's legible. If your AI agents are doing good work but you can't see it, you're flying blind — and the moment something goes sideways, your only option is to stop everything and start over.

With v1.0.165, that changes.

A Single Console for All Your Agents

We've shipped a persistent, real-time agent workspace — a live console that shows every agent working across every product you operate, all in one place. No more clicking into individual pipeline pages to check status. No more wondering whether the AI Architect has finished decomposing that feature request, or whether the Test Engineer has caught up with the latest PR.

The feed streams directly from our agent_logs infrastructure using Server-Sent Events. Activity appears as it happens: tool calls, subtask completions, decisions, outputs. The kanban alongside it groups agents by status — Running, Waiting for Review, Paused, Completed — so you get an instant read on the health of the whole platform at a glance.

Fluid Human-in-the-Loop Collaboration

Watching is useful. Being able to act on what you see is transformative.

We've added three human-in-the-loop controls that let you collaborate with any agent, at any moment, without tearing down the pipeline:

Interrupt — Stop an agent mid-task. It pauses cleanly, no work lost, and waits for you.

Inject Instructions — Send the agent a message. It picks up your input in its next reasoning step. You don't need to stop it; you don't need to restart anything. Just tell it what you want it to consider.

Take Over — Suspend the agent entirely, do the work yourself (or part of it), and hand back control. The agent resumes from wherever you left off.

These controls are deliberately fluid. Our existing approval_gates system handles structured, pre-defined checkpoints — useful for compliance sign-offs and release gates. HITL controls are for everything else: the course corrections, the knowledge injections, the edge cases that no pipeline spec could have anticipated.

Why This Matters

The most common objection to autonomous AI systems isn't capability — it's trust. People want to know they can intervene. They want to see the work happening, not just the outputs.

Devin built trust by giving developers a collaborative editor and shell where they could take over at any moment. Cursor's Cloud Agents made task progress visible through a kanban that moves in real time. We've taken both ideas and applied them to the broader problem of running an entire autonomous SaaS operation.

When you can watch 20 agents working in parallel and reach in to redirect any one of them without disrupting the rest, the platform stops feeling like a black box and starts feeling like a team.

Available Now

The agent activity feed is live for all workspaces. No migration required. Owner and Admin users have full access to interrupt, inject, and take over. All workspace members can observe the feed.

If you have pre-existing approval_gates configured, they continue to work exactly as before — HITL controls are additive.