Agents and models

An agent is a coding agent program and the model it uses. You pick both for every run, and for every agent of an experiment.

Agents

Agent	`harness`	What it is	Model provider	Offered first
Claude Code	`claude-code`	Anthropic’s command-line coding agent.	Anthropic	`claude-sonnet-5-5`
Claude Cowork	`claude-desktop`	The agent workspace of Anthropic’s Claude desktop app, for documents, research and longer tasks.	Anthropic	`claude-sonnet-5-5`
Codex	`codex`	OpenAI’s command-line coding agent.	OpenAI	`gpt-5.5`
Codex app	`codex-app`	The Codex mode of OpenAI’s ChatGPT desktop app, with its own browser, computer-use and document tools.	OpenAI	`gpt-5.5`
Cursor	`cursor`	Cursor’s command-line coding agent.	Cursor	`auto`

The harness value is what you pass to --harness in the CLI, or to agent.harness in the API. “Offered first” is the model the web app picks for you; with the API, CLI and MCP server the model is always required.

tracks runs create "Explain how this app is deployed" \
  --harness cursor --model auto \
  --repository https://github.com/alpic-ai/apps-sdk-template

Models

Each agent accepts the models of its provider that Tracks can price.

Claude Code, Claude Cowork
Codex, Codex app
Cursor

claude-fable-5-1, claude-fable-5, claude-opus-5-5, claude-opus-5, claude-opus-4-8, claude-opus-4-7, claude-opus-4-6, claude-sonnet-5-5, claude-sonnet-5, claude-sonnet-4-6, claude-haiku-5-5, claude-haiku-4-5

gpt-6-astra, gpt-6-sol, gpt-6-luna, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5.4-nano

The models Cursor offers Tracks’ account, read from Cursor and kept up to date. Use auto to let Cursor choose, or a Cursor model id. Ids can end in an effort level, such as gpt-5.6-luna-low, and in -thinking or -fast.

These lists change as providers release models. The web app always shows the current ones in its model picker. A run asking for a model its agent can’t use is refused, and the error lists the models it can use. A follow-up prompt can switch the model, never the agent: tracks runs send-prompt <id> "..." --model claude-opus-5-5.

Differences between agents

Browser. Claude Cowork and the Codex app always use their own browser and ignore the browser option. The other agents get a browser only when you ask for one.
Cursor runs its agent loop on Cursor’s servers. Tracks learns what each step cost only once Cursor reports it, so a Cursor run can go over its budget by about the last 2 or 3 model calls.
Skills are installed in each agent’s own skills folder, so every agent finds them the way it usually does.

Agent tools

Agent tools are what you give an agent on top of the repository. In an experiment, each variant picks its own set from your library; keep one with none as a baseline.

Tool	What the agent gets
Skill	Instructions, and the files they point to, that the agent loads when a task calls for them. A folder with a `SKILL.md` giving its name and description.
MCP server	The tools of a server you name, called through Tracks. Its credentials never enter the sandbox. See MCP servers.
Command-line tool	A program installed in the sandbox by its install script, before your setup script runs. The agent runs it by its name.
Browser	A headless Chromium the agent drives through tools. Pages can offer it their own tools through WebMCP.

A command-line tool is never announced to the agent: nothing in its prompt says the tool exists. The agent has to find it on its own, as it would on a real machine. That is what makes it useful to measure whether agents discover your CLI.

Add skills, MCP servers and command-line tools to your organization’s Library in the web app. To see whether agents actually used them, read agent tool use.

Get started

Concepts

Guides

Agents and models

Agents

Models

Differences between agents

Agent tools

​Agents

​Models

​Differences between agents

​Agent tools

Agents

Models

Differences between agents

Agent tools