Agent modes

Conductor lets you change how an agent approaches a task before or during a chat. These controls affect the current session, not the repository itself.

Use Claude Code, Codex, Cursor, or OpenCode when you want a coding agent in a Conductor workspace. You can run one agent by itself, or run multiple agents in separate tabs or workspaces.

For setup, billing, and provider-specific limits, see Harnesses.

Supported controls

Open the model picker to choose a Claude Code, Codex, Cursor, or OpenCode model. If Cursor cannot start, check that CURSOR_API_KEY is configured in Settings -> Harnesses -> Cursor. If OpenCode cannot start or has no models, check Settings -> Harnesses -> OpenCode. If Codex cannot start, check your OpenAI authentication or subscription setup.

Cursor sessions support Composer 2.5 in Conductor. OpenCode sessions use OpenCode models returned by your configured OpenCode providers.

Plan Mode

Plan Mode asks the agent to make a plan before editing files. Use it when the task is ambiguous, risky, or broad enough that you want to review the approach first.

Plan Mode is useful for:

• Refactors with unclear boundaries
• Migrations
• Multi-file product changes
• Debugging where the root cause is not known yet
• Work that should be split across multiple agents or workspaces

When the plan looks right, approve it or give feedback. If you exit Plan Mode, the agent can move from planning into implementation.

Fast Mode

Fast Mode prioritizes speed. Use it for narrow edits, simple fixes, and quick follow-up work.

Avoid Fast Mode when the task needs careful codebase analysis, large refactors, or high-stakes reasoning.

Thinking and reasoning controls

Some models expose thinking or reasoning controls. Higher settings give the agent more room to reason before answering, but may take longer or use more credits.

Use higher reasoning for architecture, debugging, migrations, and code review. Use lower reasoning for straightforward edits and short questions.

Codex personalities

Codex sessions can use personalities to change how Codex approaches work. Personalities are session-level controls. Use them when you want a different working style without changing repository instructions.

Codex goals

Codex sessions can work toward a goal: a standing objective that Codex keeps pursuing across turns instead of stopping after a single reply. Use a goal for longer tasks that should keep moving without a new prompt at every step.

To set a goal, type /goal followed by the objective, or click the goal toggle in the composer to send your next message as a goal.

While a goal is active, a goal bar above the composer shows the objective, its status, and its token usage when the goal has a token budget. Use the goal bar to pause, resume, edit, or clear the goal. Stopping a turn also pauses the goal, so Codex does not keep working after you ask it to stop.

Goals are available for Codex sessions in local workspaces. Cloud workspaces do not support goals yet.

Skills

Codex, Claude Code, and OpenCode can use skills in Conductor. If you already have repo or user skills for one agent, you can often reuse them so multiple agents follow the same project conventions.

Run agents together

Use separate tabs when you want Claude Code, Codex, Cursor, or OpenCode to work in the same workspace. Use separate workspaces when the work should happen on different branches.

Repository guidance

Mode controls are temporary. For durable guidance, use Repository Settings or checked-in instruction files such as AGENTS.md, CLAUDE.md, or skills.