Skip to main content
An agent is a durable coding-agent session inside a Task. The Agents panel is where agents are listed and organized. You interact with an agent through one of two interfaces: Chat or Terminal. The interface changes how you work with the agent and how much can change after it starts. It does not change what the agent is — both are agents, both live in the Agents panel, and both can be referenced, organized, and nested the same way. A plain Terminal is a different thing entirely: it is a shell session and lives under Terminals. The full distinction is covered in Terminal Agents VS Terminals.
The Agents panel was previously called Chats. Existing agents, links, and history are unchanged — only the naming moved to match the model: the agent is the durable thing, and Chat and Terminal are interfaces on it.

​
What Agents Keep

Every agent keeps:
  • conversation history or terminal transcript
  • the selected coding agent and model for each run
  • permission and thinking-effort choices
  • artifacts, file changes, tool calls, and follow-up context it produced
Agents are separate sessions. Changing settings on one agent does not rewrite another.

​
Composer

The composer is the prompt area for an agent on the Chat interface. See Composer for mentions, slash commands, drafts, quotes, and context usage.

​
Changing Settings During A Turn

  • Permissions apply to the running turn right away. See Permissions.
  • Model, thinking effort, fast mode, and provider profile apply from the next turn. If you change one of them and then steer a message into the running turn, Traycer asks End the current turn to send with new settings?. End turn & send ends the turn and resends your message with the new settings. See Message queue.
When a provider reports a usage limit, see Usage limits and model routing.

​
Permissions

Permissions control how much the agent can do without another approval. New conversations start in Full access. To change that default on all your machines, set New conversations start in under Settings › Permissions › Modes, or choose Permission settings… at the bottom of the picker. Any conversation can still change its own mode from the composer. Supervised is the safest mode, and a good choice while you learn how agents behave in your repository. A mode change during a turn applies to that turn right away. Switching to Full access also approves requests that are already waiting. Some coding agents offer fewer modes, and the picker marks the others as not supported. Amp and Cursor run only in Full access, and Devin and Hermes Agent have no Supervised mode. See Agents & Models.

​
Auto Mode

Auto is experimental. We’re still improving its reliability and speed. It appears last in the permission picker, with an Experimental label. In Auto, reads, searches, and edits run as they would in Auto-accept edits. Each command goes to a judge model first:
  • If the judge approves it, the command runs without a prompt.
  • If the judge flags it as risky, is unsure, runs out of time, or can’t run, you get the normal approval card. The card says why it was sent to you.
The judge never refuses an action on its own; anything it does not approve comes to you. The one exception is a conversation another agent started with nobody watching it: there is no one to ask, so the action is refused and the chat says so. When Traycer’s judge reviews the conversation, switching to Auto mid-turn applies right away, but approvals that are already waiting still ask you. When the coding agent’s own classifier reviews instead, Auto can’t be switched on during a running turn: the picker disables it and says why. Choose Auto before your next message, or pick Traycer’s judge on that provider’s Permissions tab in Providers to switch now. Mid-turn, Auto is also unavailable for a moment while Traycer checks which judge would review the conversation. The picker then says “Still checking which judge reviews this conversation.”, and Auto can be switched on once that’s known. The picker’s Auto option shows a short description only. Which judge reviews and how it is billed, your rules, and a log of recent decisions live in Permissions settings.

​
Approvals

When an agent needs your OK, an Approval needed card appears in the composer area with Approve and Deny. When two or more requests are waiting, the card also offers Approve all and Deny all. File edits waiting for approval get their own File edit approval card with the same controls. When the card shortens the input or shows only a summary, expand Full command (Full input for tools other than commands) to read all of it before you decide. A command keeps its original line breaks. The card leaves this out when it already shows the whole input. When Auto mode sends a command to you, the card can offer Allow from now on…. It opens the Rules tab of Permissions settings with an allow rule drafted from that action. Nothing changes until you save the rule.

​
Thinking Effort And Fast Mode

Thinking effort controls the reasoning level for models that support it. The available levels depend on the selected coding agent and model. Fast mode appears only when the selected model offers a faster service tier. Its button is named after the provider’s tier, for example Fast. It is a Chat-interface control, not a Terminal launch option. Both live in the footer of the model picker: drag the Thinking effort slider to change the effort, and click the fast-mode button to turn it on. The model chip shows the effort you picked. While the picker is open, ⌥1–9 / Alt+1–9 sets the effort and ⌥0 / Alt+0 toggles fast mode. To pick the effort from a list of buttons instead of a slider, change Reasoning control in Layout › Composer.

​
Attachments And Voice

The Attach image button attaches images only. Pasting or dropping an image on the composer attaches it too. If the selected model can’t read images, the composer says so; remove the images or switch models. To give the agent files or folders:
  • Mention them with @. See Mention context.
  • Drag a file or folder from the File Tree into the composer. It becomes an inline @ mention, just like one picked from the mention picker.
  • Paste or drop files from your computer. Traycer inserts each file’s path, relative to the workspace when it can.
Voice input turns speech into prompt text. It is useful for longer instructions, quick notes, or hands-free follow-up while reviewing code. Click the microphone in the composer, or press ⌃⇧M / Ctrl+Shift+M, to start dictating. Tap the shortcut to start and tap again to stop, or hold it to talk and release to stop. Speech is transcribed on the device; the first time, Traycer downloads the speech model. Finalized phrases are inserted as you pause. Click Stop to finish transcribing the remaining speech. Cancel or Esc discards only unfinished speech; text already inserted stays. Turn voice input off with Voice input in General settings.

​
Message Queue

Use the Message queue to reorder, edit, pause, or steer follow-ups while an agent runs.

​
Agent Questions

Answer structured questions inline or compare an alternative answer in a fork. See Agent questions.

​
Interactive Pages

An agent can show an interactive page inside its reply. While it prepares the page, the chat shows Building page. Pages also appear in the Task’s Files panel and are visible to everyone in the Task, including pages created in private chats. Hover the page to choose Expand to open it in a separate panel, or Download HTML to save an HTML copy. On a phone, open Page actions (⋯) and choose Open full screen or Download HTML. If a page stops responding, choose Reload to run it again. When an updated page includes an Updated from an earlier page link, use it to return to the earlier version.

​
Message Actions

Hover a message to show its actions. They are icons; hover one to see its name. Edit and Delete are unavailable while a turn is running. When you edit a message and that turn or a later one changed files, Traycer asks Submit from a previous message?:
  • Revert rolls files back to how they were before this message. Leave Also revert N artifacts checked to roll back artifacts too.
  • Don’t revert keeps the files as they are.
Either way, the messages after the edited one are cleared and the edited message is sent. To roll back one turn’s files without changing the conversation, see Undo a turn.

​
Forking

A fork is a new agent that starts with a copy of the conversation up to a chosen point. The original agent is not changed. Click Fork conversation under a finished reply to open Fork agent:
  • Title names the fork. It starts as “Fork - ” plus the original title.
  • Harness picks the coding agent, model, thinking effort, and fast mode. The fork can use a different coding agent from the original.
  • The host and workspace controls choose where the fork runs.
Click Fork, or press ⌘Enter / Ctrl+Enter. Where the new tile opens follows Open new tiles in Opening behavior. A fork on another host gets a copy of the transcript. Folder paths and uncommitted changes stay on the original machine, so pick that host’s workspace for the fork. Question cards have their own fork actions, Cross Question and A/B Fork. See Agent questions. To ask a quick side question without forking by hand, use /btw. See Side chats.

​
Continue A Subagent As A Chat

When a coding agent delegates work to a native subagent, its activity appears in a subagent card inside the conversation. Hover the card’s Open as chat icon to identify it, then click it to read the subagent’s conversation. This view is read-only. Back to chat returns to the parent conversation. For a Claude Code or Codex subagent in your own chat, click Continue as chat in the subagent view to carry on its work yourself. The button is disabled while the subagent is still running. The chat’s host must support this action. To create the independent chat for the first time, it must still be able to read the provider’s record of the subagent. Traycer opens an independent chat with the subagent’s conversation, nested beneath the parent in the same Task. It runs on the same host and shares the parent’s workspace, including any worktree; it does not create a separate worktree. Messages you send there do not go to the original subagent or its parent. Choosing Continue as chat again opens the chat already created for that subagent.

​
Undo A Turn

Click Undo on an assistant turn’s Changes card to restore its files and artifacts. The Files changed panel can undo everything the agent changed. See Undo a turn.

​
Work In Progress

While an agent works, panels stack above the composer: Stopping an agent that has sub-agents running asks Also stop N sub-agents?. Choose Only this agent, or Stop all to stop everything it delegated to. To fold any of these panels into a small chip, use Layout › Composer. When a reply ends with suggested next steps, each one is a button. Click it to send that suggestion as your next message, or use its copy icon to copy the text.

​
Read Replies Aloud

Use the speaker button below a reply, next to Copy, to Read reply aloud. Press Stop reading to stop playback. The button appears only when your device has an available on-device voice. Reading skips code blocks and diagrams. Only one reply plays at a time; starting another stops the first. Playback also stops when you switch away from the chat or a new reply arrives in it.

​
The Terminal Interface

An agent on the Terminal interface starts with a fixed launch setup that you choose in Start a new agent. What goes into it, and why it stays fixed, is covered in Launch Setup. Traycer stores the upstream coding-agent session id so the agent can resume the same underlying session on its original host. Agents on the Terminal interface also receive Traycer’s agent context: artifact guidance, skills, and agent-to-agent instructions. What other agents can do with them is covered in Agent-to-Agent. Not every coding agent can back the Terminal interface. The supported set is listed in Agents & Models.

​
Creating An Agent

Click + in the Agents panel header to open Start a new agent. The dialog opens in the interface you used last: this Task’s unsent new-agent draft if there is one, otherwise the interface of the Task’s most recent agent. Use Switch to Terminal or Switch to Chat above the composer to change it:
  • Chat for an in-app conversation you steer turn by turn. Write a prompt and click Send to create the agent and start its first turn.
  • Terminal for a terminal-style session with a fixed launch setup. Configure the launch and click Start.
⌘Enter / Ctrl+Enter also sends or starts. What you type in the dialog is kept as a draft until you send it. The dialog has no name field. Until the agent has a generated title, it appears as Untitled agent; use Rename in its row menu to name it yourself. To create a child agent beneath an existing one, open the agent row’s ⋯ menu, or right-click the row, and choose New child agent. Delegated work then stays attached to where it came from.

​
Hierarchy

The Agents panel is a tree, not a flat list. Top-level agents appear at the root of the panel. When an agent starts another agent, the new agent appears beneath the agent that spawned it. This creates a visible lineage:
These nested entries are often the sub-agents of the original work. They are separate agents with their own history, but the tree shows which agent created them. The delegation model is covered in Agent-to-Agent. Agents nest only under other agents, either because one agent started another or because you moved it there. A child can use either interface. To move an agent yourself:
  • drag its row onto another agent’s row to nest it there, or
  • drag it onto empty space in the panel to move it back to the root.
You can’t drop an agent onto itself or onto one of its own children. Parent and child relationships are navigation and provenance. They do not restrict which agents you can reference — see Agent-to-Agent.

​
Panel Controls

​
Header Menu

The ⋯ menu in the panel header offers: To search chat titles and messages across Tasks, use Search chats (⌘⇧F / Ctrl+Shift+F). See Search and navigation.

​
Filter Agents

The Filter agents menu controls what the tree shows and in what order. Reset view appears once anything differs from the defaults. Filtering does not change agent identity or the hierarchy. When a filter is active, Traycer keeps the relevant parent chain visible so a matching child agent does not appear detached from its context. An empty result means no agent matches the filter, not that the Task has no agents.

​
Row Menu

Open an agent row’s ⋯ menu, or right-click the row:

​
Archiving

Archive an agent to hide it and the agents nested under it. Use Archive in the row menu, or the archive button that appears when you hover an idle row.
  • You can’t archive an agent while it is working or has background work running.
  • Nested agents that are open, working, or unread stay visible.
  • To see archived agents, set Show to Archived only or All chats in Filter agents, then choose Unarchive to bring one back.

​
Row Status

An icon at the start of a row shows when the agent is running, has background work, or needs your attention, for example an approval. Agents on the Terminal interface can also show:
  • Asleep: the session is paused. It resumes on the next message or when you open the agent.
  • Stopped: the session has ended.

​
Claude Scheduled Jobs

When Claude Code schedules a job in a Chat session, open the chat’s Background panel to see its schedule and prompt. Next run appears when a future run time is available. Ask Claude to cancel a scheduled job; it has no individual Stop button in the panel. Scheduled jobs last only as long as their session; they do not survive the session ending.

​
Session State

Each agent keeps its own history and active run state. An agent on the Chat interface is flexible turn by turn. An agent on the Terminal interface is fixed after launch, because it runs through a terminal-style coding-agent session.