Agent Roles
Every built-in agent role in Cerevisor, what it's for, its smart defaults, and when to use which.
Cerevisor ships with 17 built-in roles. A role is mostly a starting point; it sets a few smart defaults (which subagent type to use, which model to prefer, which tag color the badge has), then you write the instructions that turn the role into a specific agent for your specific workflow.
You can also pick Custom and define everything from scratch, or create your own reusable agents via Settings → Agents → Custom Agents.
The full list
| Role | Tagline | Default subagent | Default model | When to pick it |
|---|---|---|---|---|
| Researcher | Investigate & discover | explore | haiku | Web research, source gathering, market scans, anything that involves crawling and synthesizing external information. |
| Architect | Design & plan | plan | sonnet | High-level design decisions, sequencing, breaking a big goal into ordered steps. |
| Developer | Build & implement | general-purpose | sonnet | Writing code, generating files, executing concrete build tasks. |
| Reviewer | Review & validate | explore | haiku | Checking another agent's output, looking for bugs, validating against criteria. |
| Documenter | Document & explain | general-purpose | haiku | Producing user docs, READMEs, API references, internal documentation. |
| Tester | Test & verify | general-purpose | sonnet | Writing or running tests, verifying invariants, integration testing. |
| PowerPoint Presentation Builder | Slide decks & narratives | (auto) | (auto) | Producing slide decks from research or planning output. Pairs well with the pptx skill. |
| Blog Post Writer | Long-form article drafting | (auto) | (auto) | Drafting 600-2000 word blog articles. Pairs well with cerevisor-content, ai-harness-content, and the writing skills. |
| Excel Expert | Spreadsheet modeling expert | (auto) | (auto) | Building or editing .xlsx financial models, data sheets, dashboards. Pairs with the xlsx skill. |
| Newsletter Writer | Audience-ready editions | (auto) | (auto) | Putting together a recurring newsletter from upstream content. |
| SEO Content Optimizer | Search visibility tuning | (auto) | (auto) | Editing existing content for search visibility — title tags, headings, keyword fit, snippet structure. |
| Meeting Notes Summarizer | Decisions & action items | (auto) | (auto) | Turning raw meeting transcripts into structured summaries with decisions and action items. |
| Research Brief Writer | Structured research briefs | (auto) | (auto) | Converting research findings into structured, decision-ready briefs. |
| Social Media Caption Writer | Caption & hook generation | (auto) | (auto) | Producing LinkedIn / X / Bluesky captions from longer-form content. |
| Coach | Integrate & guide | general-purpose | opus | A senior agent that reviews other agents' decisions, mentors the workflow, makes judgment calls. Defaults to the strongest model. |
| Situation Reporter | Summarize handoff state | plan | sonnet | Composes a structured handoff after a wave completes (default output file: wave-situation.md). Use it as a checkpoint between waves so downstream agents inherit one canonical read of what's done, what's missing, and what's still unknown. Pairs with the Situational Awareness pillar. |
| Custom | Define your own | auto | auto | Everything starts blank. You set role description, instructions, model, tools. |
"Auto" means what?
For subagent type or model preference, auto means: at run time, the orchestrator picks a sensible default based on the agent's instructions, the provider you're using, and the model catalog available. You can always override per agent.
Subagent types
The four subagent types influence how the harness scopes the agent's permissions and which tools it gets by default.
| Subagent type | Tool defaults | Best for |
|---|---|---|
| explore | Read-only tools (read, grep, glob, web_search, web_fetch). No write or execute. | Researchers, reviewers — agents that should look at things but not modify the world. |
| plan | Read tools + thinking modes that produce structured plans. | Architects, planning agents. |
| bash | Bash execution + file read. No file write. | Agents that run shell commands but shouldn't write files directly. |
| general-purpose | Full toolset (read, write, edit, bash, web). | Developers, testers, coaches — agents that actually do things. |
You can override the default tool set per agent in the Agent Config popup → Allowed tools.
Model preferences
When the role's default model preference is set (e.g. haiku for Researcher), the orchestrator chooses the cheapest capable Haiku-class model on your current provider. Override per agent via the Model picker in Agent Config.
| Preference | What the router picks |
|---|---|
| haiku | The cheapest Haiku-class model available on the resolved provider. On Anthropic: claude-haiku-4-5. On Ollama: the smallest model you have pulled, or one you've tagged as "fast". |
| sonnet | Mid-tier — Sonnet on Anthropic, a 7-13B model on local providers. |
| opus | The most capable model on the resolved provider — Opus on Anthropic, the largest model you've configured locally. |
| auto | The router's call. Currently maps to sonnet unless overridden by your provider's catalog. |
The router also enforces context-window and output-token caps per model, see Reference → Settings → Providers for the model catalog mechanics.
Creating a Custom Agent
Two ways:
- Inline: pick Custom as the role when adding the agent. Configure everything in the Agent Config popup.
- Saved template: open Settings → Agents → Custom Agents → New. Fill out the template (name, role description, default skills, default model). Save. The template now appears in the agent palette and you can drag it onto any canvas.
Saved custom agent templates sync across all your Cerevisor workflows.
Reordering and hiding roles
Don't use the PowerPoint role? Hide it.
Settings → Agents → Roles. Drag to reorder the palette. Toggle visibility per role. Reordering affects the 0–9 placement shortcuts (1 = first visible role, 2 = second, etc.).