Agents Overview
7 specialized AI team members. Each with a role, personality, and rules.
The Team
Every agent has a name, an avatar, a job title, and a set of skills they are allowed to use. They cannot operate outside their scope. This is how you get reliability without micromanagement.
How Agents Work
Each agent has scoped skills, a personality defined in soul.md, and operational rules defined in AGENT.md. They can only use the skills assigned to them. Cortex gets all 14 workflow skills because it orchestrates. Loom only gets copywriting and content skills. Signal only gets social and distribution skills.
This scoping is what makes the system reliable. An agent that cannot access a skill cannot misuse it. When an agent hits the boundary of what it can do, it escalates to the human or hands off to another agent.
Agent Workspace
Every agent has a workspace folder at .claude/agents/{name}/ containing four files.
| File | Purpose |
|---|---|
AGENT.md | Name, description, model, scoped skills, rules, inputs/outputs, hard bans, and escalation triggers. |
soul.md | Voice layer. Tone examples, operating principles, banned words and phrases. |
identity.md | Emoji, role title, and a one-liner description. |
HANDOFF.md | Live state. Current task, open loops, what the agent is waiting on, and the next action. |
Scheduled vs Interactive
Agents run in two modes. Interactive mode means the human is present, typing commands, reviewing output in real time. Scheduled mode means the agent runs unattended on a cron-like trigger.
Only two agents run on schedule: Cortex and Radar. They execute 3 times per day via RemoteTrigger API. Cortex plays 5 scheduled roles: Scout, Task Proposals, Chronicle (nightly journal + memory), Mirror (weekly reflection), and Compass (monthly reflection). Radar handles outbound work like repo monitoring and signal detection.
Scheduled agents create branches, commit, push, and open PRs. They never push directly to main. The human merges when ready.