Specialists, Skills, And Models

These dashboard surfaces work together even though they live on different pages.

Agirunner treats agent behavior as a combination of identity, instruction, and model routing rather than as one giant prompt blob. That separation makes the system easier to tune and easier to explain.

Specialists

The Specialists page defines agent identity:

• prompt baseline
• tool grants
• linked MCP surfaces
• execution environment
• skill assignment
• active and inactive posture

In the dashboard implementation, specialists are durable role definitions with list, edit, duplicate, delete, and active-state controls. That is a good sign: these are long-lived product objects, not temporary prompt snippets.

This is also where several other dashboard surfaces come together. A specialist is the point where:

• Models decide which provider or model route to use
• Tools decide which built-in capabilities are allowed
• MCP Servers decide which external tool surfaces can be attached
• Environments decide where the specialist’s work actually runs
• Skills and shared instructions shape reusable execution guidance

Skills

The Skills page is the shared instruction library that specialists can reuse.

Skills separate reusable guidance from role identity. That keeps role definitions smaller and makes instruction content portable across specialists.

This is also where imported community content lands after a playbook import. A playbook pulled from agirunner-playbooks can create tenant-local specialist and skill copies here for later inspection and editing.

Models

The Models page manages provider connections, discovered catalogs, system defaults, and role-specific assignments.

That surface is deeper than “pick a model.” It is where operators shape the routing posture the platform will later hand to the runtime as an explicit contract.

Why The Split Is Useful

This separation lets teams answer different questions independently:

• who is this specialist supposed to be
• what reusable guidance should it inherit
• what model or provider should power it

That is a much cleaner operating model than hiding all three inside one prompt string.

It also means poor model choices stay visible instead of getting buried inside prompt text. For real specialist work, Agirunner currently performs best when these roles use reasoning-capable models. The recommended default is gpt-5.4 with at least low reasoning, and usually medium, rather than low-capability or non-reasoning models.

How It Connects To Execution

The specialist page is not just an authoring surface. It is where the platform assembles the contract the runtime will later enforce for a claimed task.

When a workflow creates specialist work:

1. the workflow and playbook determine what kind of specialist is needed
2. the specialist definition contributes prompt identity, skills, tools, linked MCP servers, and environment choice
3. the platform resolves model and policy posture
4. the runtime claims the task and exposes only the resulting allowed execution surface