Agent mode
Persistent teammates with a role, rules, an engine, allowed folders and long-term memory — with conversation tabs, task modes and handoffs between teammates.
Code mode gives you terminals. Agent mode gives you teammates: agents that keep a role and a memory across conversations, and that you talk to in a structured conversation view instead of a raw terminal. Open it with Ctrl+Alt+1.
The left column lists your teammates; the main area shows the selected one with its conversations.
Create a teammate
You have two ways to start:
- Create with AI — describe the teammate you need in a few sentences, pick the project it works in, and click Generate draft. Presets get you started: QA & E2E test lead, Security & CVE reviewer, Frontend design specialist and Cloud DevOps engineer. You always review and adjust the draft before it is saved; workspace edits start with guided approval.
- Fill in yourself — opens the teammate settings directly.
Teammate settings
| Field | Purpose |
|---|---|
| Name and Role | How it introduces itself and what it is responsible for — for example Mira, QA lead |
| Description | What this teammate is for, in one or two sentences |
| Brief | Standing context about the project and your expectations |
| Working rules | Things it must or must not do — Never push. Never weaken tests. Report briefly. |
| Engine | Claude Code or Codex |
| Model | Leave empty for the CLI’s default model |
| Effort | Low, Medium, High or Max |
| Authority in the project | Read only, Guided (it plans first, you approve the edits) or Edit |
| Places | The project folders it may work in. The first one is its home; every folder you add is approved for agents |
| May hand off to | Other teammates it can pass subtasks to |
| Long-term memory | Whether saved facts are added to every conversation |
Name, role, description, brief and working rules are required. While you fill them in, the brief preview on the right shows exactly what the agent will receive at the start of every conversation — identity, places, access and rules.
A teammate without a project folder can talk, but not work in code.
Conversations
The header shows the teammate’s name and role, whether its engine is installed, its home folder (click to show it in Explorer), its authority and whether memory is on.
Each teammate has tabs — one per conversation — so you can keep several threads with the same teammate apart. New conversation starts a fresh one; double-click a tab to rename it, right-click to delete it. The 30 most recent conversations are shown.
Claude Code and Codex stream their work into the conversation as it happens: every step, tool call and command is listed and can be expanded, and each answer shows the engine, tokens, cost where reported and duration. The composer works like the one in Chat: attachments, @ files, dictation, prompt enhancement and effort.
Task modes
Above the conversation you choose what kind of work the next turn is:
| Mode | What the teammate does |
|---|---|
| Complete | Works on the task with its configured authority |
| Review | Reads the relevant code and recent changes and reports findings ranked by severity, with file and line. Changes nothing |
| Harden | Runs the project’s existing checks — tests, type checks, linters — and reports. Only fixes failures that are clearly in scope, never weakens or skips tests |
| Explore | Explores two or three distinct options with trade-offs and a recommendation. Changes nothing |
Review and Explore are always read-only, whatever the teammate’s authority.
Guided turns
With Guided authority, the teammate first answers with a plan. Under the plan you get Approve & implement, Change plan and Reject. Only after you approve does it edit files — and before it does, PaneScale takes a checkpoint. Below the result you’ll find Changes to review the diff and Rewind step to undo it exactly.
Memory
Open Memory in the teammate header to see what it remembers. A memory is a short fact with a key — Styling guideline, Release process — and content.
- Add memory writes one by hand. Click a memory to edit it; search filters the list.
- Learn from this conversation reads the open conversation and proposes facts worth keeping.
- Reflect now reads all of the teammate’s conversations and proposes durable facts.
- Proposals are listed with checkboxes — keep what is right, Save selected.
- Use memory switches whether memories are added to conversations.
Memories are added to future prompts as context, never as instructions, so a memory can’t override your rules. Project-wide skills work the same way.
Handoff between teammates
When a subtask belongs to someone else, click Hand off in the teammate header:
- Pick the receiving teammate — only those listed under May hand off to are offered.
- Describe the subtask.
- Choose whether to include the latest messages as context.
- Click Send handoff.
The receiving teammate gets a clearly labeled handoff message in a new conversation and works with its own authority and rules. To hand a terminal session to another agent, use the pane menu instead.
Delete a teammate
Use the menu next to a teammate in the list and choose Delete teammate. Its settings and memory are deleted; its conversations stay in the database and can still be found in the command palette.