Agents and models

Each task picks an agent, the coding tool that does the work, and a model, which drives it. The agent stays the same for the whole conversation; the model and reasoning level can change on every turn.

Agents

Agent	Steer a running turn	Asks you questions	Subagents	Strongest reasoning
Claude Code	Yes	Yes, and plan approvals	Yes	max
Codex	Yes	Yes	Yes	max
OpenCode	—	Yes	Yes	max
Kilo Code	—	Yes	Yes	max
pi	—	—	—	high

Every agent runs as its own command-line tool, with its own session, inside the conversation’s sandbox, and every agent shows its work in the same Chat and Trajectory views. All five can use your MCP servers, skills and plugins.

• Steer a running turn: a message sent with Steer now reaches the agent before its next tool call. With the other agents, a message sent during a turn waits for the next one. See Follow-ups.

• Subagents: when an agent delegates work, the Tasks and subagents window shows each subagent’s assignment, model and reply, and you can open the subagent’s own conversation and steps. See Tasks and subagents.

Models

The composer lists the models enabled for agent tasks on your account, which during the beta can be a subset of the full model catalog. A new task starts on the default model. The control beside the message box starts each follow-up on the model the previous turn used.

Reasoning

Where the agent and model support it, the Reasoning control sets how much the model thinks before acting. The levels each agent offers differ; higher levels are slower and use more of your quota.

Choosing

Work	A good starting point
Mechanical sweeps, renames, small fixes	A fast model, low or medium reasoning
A bug with a clear reproduction, a focused feature	The default model, high reasoning
Open-ended design, hard debugging	A larger model, the strongest reasoning

If you already use one of these agents locally, pick the same one here: it reads the same AGENTS.md or CLAUDE.md and behaves the way you know.