Agents
The planner, the coder and the reviewer, which model each uses, and what each may touch.
A run is carried out by several agents, each a Claude model with its own instructions and tools.
| Agent | Does | Model setting |
|---|---|---|
| Planner | Triage, the plan, and plan chat. Reads code, the code graph, memories, skills and connections; never writes. | Planning & reviews |
| Coder | Writes the change, runs the build and tests in the sandbox, fixes CI failures and answers feedback. | Writing code |
| Reviewer | Reviews CoderHelm's own pull requests and your team's (see The PR reviewer). | Planning & reviews |
| Writer | Pull request descriptions, replies to comments, and release notes. | Planning & reviews |
Choose the models on AI Models (admins only). The models on offer are Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 4.5, plus older models for teams that pinned one. A model that Anthropic retires runs as its replacement automatically.
What agents can touch
- Your code: read through the GitHub App. Changes are pushed only to
coderhelm/branches and arrive as pull requests. - The sandbox: builds, linters and tests run in an isolated build environment with no credentials and no access to your systems.
- Your data: only through data sources and team tools, which are read-only.
- Outside tools: only through connections an admin switched on.
Team agents
Your team can add its own agents, written as Markdown files in a skill repo. A team agent is a specialist the main agents can ask, for example "the payments expert". Team agents answer questions with read-only tools; they never write code. See Skills.
Instructions
Besides skills, admins can set plain-text instructions in Settings, for the whole team or one repository: general instructions, rules, the voice for PR descriptions and comments, and git conventions (branch, commit and pull request title templates).