The agent workspace
Open any agent from My Team and you land in its workspace — a single screen that shows what the agent is doing right now, lets you steer it, and exposes everything it knows, holds, and is allowed to do.
The workspace has three layers: a control bar at the top, the coaching panel just below it, and six tabs — Live, Tasks, Knowledge, Tools, Permissions, and Health.
The control bar#
Across the top you'll find the agent's identity and its live controls:
- Instruct — opens the command drawer to send the agent a direct instruction (it lands in the agent's inbox as guidance). See Chatting with agents.
- Pause / Resume — stops or restarts the agent's work.
- Take Over / Release Control — puts a human in direct control. While you're in control, gated actions execute as your decision and are audited as such. Click again to release the agent back to autonomous running.
- Mode — the supervision-mode picker (Inherit / Strict / Balanced / Autonomous), with a Pinned chip when the mode is set on this agent rather than inherited from the workspace. This is covered in depth in Supervision & modes.
The coaching panel — "Improve This Agent"#
Right under the control bar is a highlighted panel: Improve This Agent — Coach your AI employee, help it perform better in your business. It has two buttons:
- Teach something — opens the teaching modal, where you record durable guidance in one of four forms:
| Teach type | Prompt | Use it to |
|---|---|---|
| Knowledge | "What should I remember?" | Tell the agent something important about people, work, or preferences |
| Behavior | "Set Behavior" | Define how the agent should communicate or work |
| Boundary | "Set Boundary" | Set limits on what the agent can do |
| Autonomy | "Set Decision Authority" | Decide when it should act on its own vs. ask you first |
- Give feedback — opens the same command drawer as Instruct for a quick correction like "Follow up after 2 days instead of 1" or "Never send discounts."
Coaching is how an agent improves over time. What you teach here becomes durable memory and rules that show up in the Knowledge tab. See Memory & learning.
Live#
The default tab — the agent's current brain.
- Current Action. A "Now" panel showing what the agent is doing this moment: the task title, its tone and goal, live progress, and the last update time. When there's no live telemetry yet, it says so plainly rather than inventing activity.
- Activity heatmap and streaks. A calendar heatmap of actions over time, with longest streak, average per day/week, total actions, 7-day success rate, and time saved.
- Upcoming / next actions. The next few things the agent intends to do, each with a when and the rule that scheduled it.
- Needs Your Decision. If the agent has parked an action for approval, it surfaces here with the consequence ("After approval: …"), risk, confidence, and the reason it needs you. You can act on it without leaving the tab.
- Reasoning stream. A running log of the agent's thinking, grouped into four lanes:
| Lane | Shows |
|---|---|
| Learned | New things the agent picked up |
| Decisions | Choices it made, including approval outcomes |
| Actions | Things it did — tool calls, messages, scheduled runs |
| Waiting | Where it paused, and what it's waiting on |
Tasks#
The agent's work queue. A badge on the tab counts pending items.
- Waiting for Approval — actions the agent has drafted and parked for your sign-off. Approve or reject inline.
- Pending Actions — toggle between Active and History. Active tasks show a humanized schedule ("Scheduled in 2h · 3:00 PM", "was due 3h ago"). Cancellable tasks have a Cancel button that prompts for an optional reason.
- Decision History — the last resolved approvals, each tagged approved or rejected, with who decided and why.
Knowledge#
Everything the agent remembers and the rules it follows — the editable memory built up through coaching and observation. It's organized into:
- What the agent knows — observed facts and preferences (e.g. "Prefers email over calls"), each with a caption noting where it came from.
- Behaviors — how the agent communicates and works.
- Boundaries — hard limits on what it may do.
Click any item to edit or delete it. Use Teach something in the coaching panel to add new entries.
Tools#
The connected systems this agent can reach — its senses and hands. Each tool card shows:
- The tool's brand logo, the connected account, and a Connected badge.
- What it can do — either the granted OAuth scope or a capability list.
- A Permissions dial — the per-tool control that decides how the agent may use this tool: Off, Read-only, Asks first, or Autonomous (a Custom state appears for advanced combinations). A one-line preview spells out what a write with this tool would do right now.
- Detach — removes the tool from this agent.
At the bottom, Connect more apps takes you to Connecting tools. The dial itself is documented in full in Tool permissions.
Permissions#
The governance surface for this agent. It has three parts:
- Supervision mode — the same Inherit / Strict / Balanced / Autonomous picker as the control bar, with a Pinned on this agent or Inherited from workspace chip and a one-click Reset to workspace default.
- Test an action — a dry-run simulator (below).
- Everything is read-only for Operator and Viewer roles, with a "changing permissions requires an Admin or Owner role" note.
See Supervision & modes and The three planes for the full model.
Test an action#
Pick a connected tool and type an action (for example GMAIL_SEND_EMAIL), then Preview decision. The simulator dry-runs the gate and tells you exactly what would happen — allow, approval, or block — and which level decided (workspace, agent, or tool). Nothing actually runs; it's a safe way to check your policy before trusting the agent with it.
Health#
The agent's fitness report.
- Work Quality — a headline score from approval outcomes over the last 7 days (Good / Ok / Needs attention, or "No data yet").
- Metric grid — reply effectiveness, corrections needed (rejected approvals), waiting for you (pending decisions), safety behavior, AI usage cost, and responsiveness (average time from you to the agent).
- Health Status — whether the agent is active (with last heartbeat), quality-of-work basis, and anything needing attention.
- Diagnostic Logs — a technical log stream for troubleshooting.
If an agent shows Needs attention on My Team, the Health tab is where you find out why — a runtime error, a stall, or pending decisions.
What's next#
- Supervision & modes — set how independent this agent is.
- Tool permissions — tune what each tool may do.
- Approvals — clear the decisions your agent parks for you.