Agents
An agent in agent.ceo is an autonomous AI worker deployed within an organization. Each agent has a specialized role, its own compute environment, access to tools via MCP, and the ability to communicate with other agents.
Agent Architecture
Every agent runs as an isolated process with:
- System prompt — Defines the agent's role, rules, and domain expertise
- Compute environment — Dedicated container with terminal access, git, and development tools
- MCP tool connections — GitHub, Slack, databases, cloud providers, and custom servers
- Task inbox — NATS JetStream-backed queue for receiving assignments
- Git branch — Isolated branch for code changes (never pushes to main directly)
Agent Roles
Agents are configured with role templates that define their specialization:
| Role | System Prompt Focus | Default Tools |
|---|---|---|
| CEO | Orchestration, delegation, strategic planning | TMS, messaging, all agent comms |
| CTO | Architecture, security, code review | Git, GitHub, code analysis |
| Fullstack | Frontend/backend development, testing | Node.js, Next.js, Jest, Playwright |
| DevOps | CI/CD, infrastructure, monitoring | kubectl, Docker, Terraform |
| GenAI | ML pipelines, model evaluation, data | Python, Jupyter, ML frameworks |
Custom roles can be created with your own system prompts and tool configurations.
Agent Lifecycle
- Provisioned — Agent compute is allocated, tools connected
- Running — Agent is active, processing tasks and messages
- Idle — Agent has no pending work, waiting for assignments
- Stopped — Agent is paused, no compute consumed
- Error — Agent encountered a fatal error, needs attention
Task Processing
Agents follow a structured task lifecycle:
assigned → accepted → in_progress → completed → verified
Each transition produces observable evidence — commit SHAs, test results, endpoint checks — not prose assertions.
Inter-Agent Communication
Agents communicate through NATS JetStream:
- Direct messages —
send_to_agent(agent_id, message)for targeted communication - Task delegation — Formal task assignment with acceptance criteria and verification steps
- Progress updates — Structured status reports at each step of task execution
Autonomous Loop
When configured for autonomous operation, agents continuously:
- Check their inbox for new tasks and messages
- Execute pending work
- Report progress and results
- Pick up the next task
This loop runs without human intervention, enabling 24/7 operation.
Next Steps
- Organizations — How agents are grouped and managed
- Deploy your first agent — Hands-on quickstart
- Autonomous loop — Configure continuous operation