Copilot for operators

Copilot is the interactive assistant in Studio’s operator workspace. Use it when a support case needs hands-on investigation, a specialist decision, or an action through local tools. It can reason over conversations, active tabs, hosts, files, procedures, memories, connectors, browser sessions, remote desktops, calls, and generated work. Its tools can inspect systems, open operational surfaces, and—after the required approval—change external state. AI support agents carry configured cases through the cloud workflow. Copilot supports the person working a case or responding to an agent’s human task. Return the verified findings to the task or support system so the support loop can continue.

​

Modes and Autopilot

Response mode and autonomy are separate controls.

​

Models and context windows

Auto chooses a balanced or higher-capability model per message. The model selector includes supported Claude models through Amazon Bedrock and OpenAI models through Studio’s authenticated proxy. Available choices and defaults depend on your organization’s model policy. You can select a different model for a conversation without changing your personal default when policy allows it. An administrator can restrict the model catalog or manage the selection for everyone; a disabled selector is not fixed by using /model. For models that support selectable windows, Settings offers:

• 200K (recommended) — summarizes sooner and keeps long conversations less expensive.
• 1M (extended) — retains substantially more context before summarizing, with materially higher cost for long sessions.

OpenAI models currently use Studio’s 200K effective context limit. The Bedrock region setting applies only to Bedrock-backed workflows; it does not change the OpenAI route. See AI provider and data flow. Changing your personal default affects new chats and standalone workflows. An existing chat with an explicit model selection keeps it unless organization policy overrides the selection.

​

Context hierarchy

Copilot’s answer can be shaped by several layers:

1. Organization context and organization tool policy.
2. Your member context and preloaded tools.
3. The active channel’s grounding and tool policy.
4. Conversation history, model, and mode.
5. Explicit mentions, attachments, and active-tab context.

Explicit context is the most reliable. Mention or attach the exact object instead of relying on Copilot to infer which host, file, tenant, or procedure you mean.

​

What you can attach or mention

• Hosts and their configured protocols.
• Files, generated artifacts, dashboard apps, and session replays.
• Procedures and procedure runs.
• Connectors, MCP servers, and tools.
• The active terminal, browser, remote desktop, dashboard, or editor.
• Selected terminal output or an image.
• Computer Use when a local desktop application is the target.

Never attach a raw secret. Reference a Key Chain entry or complete the integration’s account flow instead.

​

Ask Org Brain

When shared history matters, ask Copilot to search Org Brain before answering:

• Search organization knowledge returns source excerpts and citations for inspection.
• Ask Org Brain produces a cited answer from authorized evidence when the organization’s synthesis controls allow it.

For example: “Find the decision behind this site’s routing exception, cite the source, and separate the recorded decision from current device state.” Retrieval follows organization and source visibility. A source citation is evidence to inspect, not a new permission to open or share the underlying record. If a source has changed or access is no longer allowed, the citation may be unavailable. If synthesis is disabled or unavailable, report that limit; do not substitute raw search to bypass it.

​

Tools and approvals

Copilot tools span host access, local diagnostics, terminals, RDP, browser, Computer Use, inventory, diagrams, files, dashboards, calls, procedures, memory, connectors, MCP, planning, scheduling, and sub-agents. Studio evaluates a proposed action before dispatch: An approval is scoped to the displayed action. Allow for session is broader; use it only after you have verified the recurring target and operation.

​

Browser AI Grounding

Enable Browser AI Grounding from the composer when Copilot needs stronger visual understanding of an unfamiliar or complex webpage. It can improve browser decisions, but adds token use to browser actions. See Browser and Computer Use for takeover and browser autopilot.

​

Queue and steer

While Copilot is working, a new message can either:

• Queue — preserve the current run and deliver the message at its next completion boundary.
• Steer — inject the message into the current run so Copilot changes direction.

Choose the default under Settings → AI → Follow-up behavior. Use Command+Enter on macOS or Control+Enter for the opposite action. Steer when the target, assumption, or scope is wrong. Queue when you want the current task to finish before a related follow-up begins.

​

Slash commands

System commands change the current conversation without a long prompt. Tool shortcuts route a prompt toward a specific surface: /dashboard, /host, /hosts, /terminal, /exec, /explain, /inventory, /connector, /mcp, /procedure, /memory, /history, /video, /sip, /delegate, and /tools. Connected and dynamically discovered tools can add more shortcuts. Type / and use the menu instead of memorizing the full catalog.

​

Plans and sub-agents

Planning mode gathers evidence and renders an actionable plan. Review its targets, prerequisites, approval points, validation, and rollback before execution. For work that can proceed independently, Copilot can delegate to sub-agents. Their live progress and approvals remain attached to the parent conversation. Delegation does not lift organization, channel, member, or approval restrictions. Digital workers are different: they run in the cloud under a shared channel’s policy, with their own budgets, lifecycle, and human gates. Review tasks in the desktop Inbox or browser Work area. See Digital workers before turning an interactive workflow into unattended automation.

​

Durable outcomes

Ask Copilot to preserve work in the right form: Organization memory stores last-known work-product state, not live proof. Verify a path, process, deployment, or URL before claiming it is currently available. Keep personal observations private until they are appropriate to share. Saving a shared fact is a write, separate from read-only Org Brain retrieval, and follows the applicable approval and visibility controls.

​

Usage and cost

The transcript can show turn, token, model, and cost context. Open Usage → Per-chat for your detailed conversation history and Usage → AI for aggregate and model views. Reduce unnecessary cost by attaching precise context, compacting long chats, avoiding repeated failed tools, and using Browser AI Grounding only when the page needs it.

​

Safe operating pattern

1. Confirm organization, channel, host, tenant, and credential scope.
2. Start in Ask for explanation or Planning for a change.
3. Attach the exact active evidence.
4. Review the plan and remove out-of-scope actions.
5. Move to Default when you are ready to approve execution.
6. Steer immediately if Copilot follows a wrong assumption.
7. Validate the external result with a read-only check.
8. Save the durable outcome and set the conversation status.

​

Related