Agents And Skills

​

The split

There are three different things in the product, and they should stay separate.

​

Atlas agents

An Atlas agent is the actual meeting participant. Each agent includes:

• a realtime wrapper for live voice behavior
• a brain for heavier reasoning and backend work
• attached skills
• an allowed subset of gateway actions

​

Atlas skills

Skills live on the Atlas side. They are reusable markdown instruction packs that teach an agent how to do something well. Examples:

• meeting notes
• research synthesis
• workflow operation
• structured evaluation

Skills are not customer tools. They are behavior and context.

​

Gateway actions

Gateway actions live on the customer side. They are the actual callable tools Atlas can use through the customer gateway. Examples:

• fetch_internal_context
• trigger_followup_workflow
• lookup_candidate_packet
• query_finance_agent

​

Runtime flow

When a meeting starts, the worker loads:

• the agent config from Atlas
• the attached skill metadata from Atlas
• the gateway manifest from the customer gateway

Then Atlas does two different things:

• it exposes allowed gateway actions to the brain as callable tools
• it exposes attached skill metadata to both realtime and brain, and either side can load the full skill on demand

​

How skills are loaded

Atlas uses progressive disclosure for skills.

• the worker injects skill metadata first
• the full SKILL.md body is not injected up front
• realtime or brain can call load_attached_skill(...) when a skill is clearly relevant

This keeps context cleaner and gives Atlas a trace of which skills were actually used.

​

What the dashboard controls

The dashboard has three separate surfaces:

• Gateway
  • register the workspace gateway
  • validate the manifest and invoke path
  • inspect available gateway actions
• Skills
  • create and edit workspace skills
  • store the markdown bodies Atlas can load later
• Agents
  • edit the agent’s base prompts
  • attach skills
  • choose which gateway actions that agent can actually see

​

Hard vs soft control

Use both.

• gateway action scope is the hard boundary
  • if an agent should not use a tool, remove it from that agent’s visible action set
• skills and prompts are soft guidance
  • they shape how the agent behaves and when it uses the tools it can see

​

Setup order

For first-time setup, use Dashboard Setup. This page is about the split between agents, skills, and gateway actions.