Add and configure agents

Adding an agent

Three ways. Pick whichever feels fastest in the moment.

1. Drag from the Agent Palette

Open the palette from the left edge of the canvas (the icon strip). Drag any role card onto the canvas. Release where you want the agent. Cerevisor places it in the nearest column.

2. Right-click the canvas

Right-click any empty area. Choose Add Agent → [Role]. The agent appears at your cursor position.

3. Press Shift+1–9 (or Shift+0)

Pressing Shift plus a number key enters placement mode for the N-th agent in your palette order (built-in roles and custom templates share one order). Your cursor becomes a ghost of the agent. Click anywhere to place. Press Escape (or right-click) to cancel; pressing the same combination again also exits.

This is the fastest path once you know your palette order. Reorder in Settings → Agents to put your favorites at Shift+1 and Shift+2. (Bare number keys without Shift switch between workflows instead.)

4. Drag a Custom Agent template

If you've saved a custom agent template (see Reference → Agent roles → Custom Agent), drag the template card from the palette onto the canvas. It carries its saved skills, model, and instructions.

Configuring an agent

Click the agent's name or role badge to open the Agent Config popup. The fields below appear in that popup, top to bottom. The block under the Structured orchestration divider (Agent Type, Provider, Model, Thinking mode, Allowed Tools, On Failure, Context Notes) only appears when the workflow has structured orchestration / native mode enabled.

Name

The agent's display name on its card. Plain text. Choose something specific, "Senior researcher" beats "Researcher 2" for any workflow with more than one agent of the same role.

Role

A grid of role buttons. Click another role to change the agent's role in place. Changing the role re-applies that role's smart defaults.

Description

A textarea. Free-form text describing what kind of agent this is. The default is the role's stock description; edit it to add specifics. Example:

A researcher specialized in EU funding programs (Horizon Europe, EIC, Eurostars). Knows the eligibility quirks and reporting requirements. Cross-checks call deadlines on the official Funding & Tenders portal.

Goes into the agent's system prompt.

Instructions

A textarea. The most important field. This is where you tell the agent what to do in this specific workflow.

Good instructions:

• Are concrete about the goal.
• Specify the input the agent will receive (from upstream agents or workflow input).
• Specify the output the agent should produce.
• Mention any constraints (length, format, sources).
• Mention any tools the agent should prefer (e.g. "use the recency-research skill").

Bad instructions: "Do research." (Too vague: the agent will produce generic output.)

Output

A prose text field, plus optional structured fields under the Advanced: enforceable structure toggle.

The prose describes what the output should look like in natural language:

A 600-word markdown blog post with a title, a one-sentence summary line at the top, and 3-4 H2 sections. Tone: matter-of-fact, second-person, no hype. Include inline source citations.

Optional structured fields:

• Required headings: chip picker. Agent's output must include these H2 sections.
• Minimum words: integer. Agent's output must hit this length.
• Deliverables: chip picker. Specific items the agent must produce (e.g. "executive summary", "data table", "next steps").

These structured fields drive Cerevisor's output validation. If the agent's output doesn't include a required heading, the harness re-runs the agent with the missing item flagged.

Equipped Skills

A chip list of the agent's skills, with a picker below it to add more. Up to 6 skills per agent. Click the × on a chip to remove a skill.

On the agent's card, skills show as hexagonal slots — you can also drag skills from the Skills panel straight onto the card.

See Working with Skills → Browse and assign skills.

Provider & model

Two fields: a Provider picker and a Model selector.

By default, the agent inherits the workflow's default provider — the Inherit option. Pick a specific provider to override for this agent only (dragging a card from the Providers dock onto the agent does the same thing). Changing the provider clears any pinned model, since model catalogs are provider-specific.

The model selector lists models available on the resolved provider. Pick auto to let the model router decide based on your role + agent type + provider's catalog.

See Providers → Per-agent overrides.

Thinking mode (if model supports)

Toggles extended thinking on models that support it (Anthropic Opus 4+, Sonnet 4+, and Haiku 4.5+ via the anthropic-budget surface, plus Gemini reasoning models via thinkingBudget, and always-on reasoners like DeepSeek R1, Qwen QwQ-32B, and *-thinking variants on OpenAI-compatible endpoints). Off by default; flip on for complex reasoning tasks where the latency cost is worth it.

Agent Type

A button grid: auto / explore / plan / bash / general-purpose.

Drives the default toolset and the agent's permission profile. See Reference → Agent roles for what each type does.

Allowed Tools

A grid of tool chips covering every built-in tool: Read, Write, Edit, Bash, Glob, Grep, Search, Fetch, Find Skills, Todo Write, Todo Read, Ask User, Skill, Subagent, List Dir, Workflow Meta, Read Bash Out, Kill Bash, Browser, Remember, and the Outlook tools (Mail, Search, Get, Calendar, Event, Draft). Tools blocked by workflow permissions appear struck through and disabled.

Defaults follow the agent type. Override per agent, e.g. you might give a "Researcher" agent Search but explicitly remove Bash.

Search Provider

A button grid: Default (workflow setting) / Tavily / Perplexity / Linkup.

Picks which backend the web_search tool uses for this agent. Default inherits from the workflow setting. Options with no key configured are disabled until you add a key in Settings → Web Search.

Input files & output files

Drag markdown files from the Markdown panel into the INPUT (left) or OUTPUT (right) drop zone. In the popup, each file also carries a one-line purpose, input files get an Auto-read option and an + Image button, and output files can carry an advisory contract. See Attach markdown files.

On Failure

A button grid: retry-once / skip-and-continue / ask-user / use-default.

What the orchestrator does if this agent fails. For agents whose output is optional, prefer skip-and-continue. For agents whose failure should pause the workflow, use ask-user.

Once a strategy is selected, a fallback instructions textarea appears below the grid: extra guidance for the failure path — e.g. what to pass downstream as a placeholder.

Context Notes

Textarea. Free-form notes that get injected into the agent's system prompt as additional context. Use for one-off "remember that the user prefers dollars over euros" guidance you don't want in your global memory.

Save as Template

Bottom-left button. Saves the agent's current config as a reusable Custom Agent template. The template appears in the palette and in Settings → Agents → Custom Agents.

Delete Agent

Bottom-left button (next to Save as Template). Removes the agent and any connections that touched it. Confirms first.

Done

Bottom-right. Closes the popup and flushes all pending edits.

Editing multiple agents at once

Select multiple agents (Shift-click or drag a marquee). Most agent operations (Delete, Duplicate, Move to Wave) act on the whole selection.

You cannot open the Agent Config popup with multiple agents selected, config is per-agent.

Smart defaults that aren't hidden

Cerevisor auto-fills these based on the role you picked:

• Subagent type: see the table in Reference → Agent roles.
• Accent color: used for the role badge and connection edges.
• Tools: defaults follow subagent type.

Roles deliberately have no default model — model choice is dynamic (the router picks unless you pin one).

Anything you change in the Agent Config popup overrides these. The defaults exist so a one-click "Add Researcher" agent is useful immediately without configuration.