Agents

Agents are the core execution units in Miteos. Each agent is a specialized AI worker with access to tools, memory, and the ability to take real-world actions.

Agent Types

Type	Specialization	Primary Tools
Researcher	Find and extract information from the web	Browser, search, data extraction
Builder	Create code, files, dashboards, applications	Code execution, file system, deploy
Analyst	Process data, generate insights and reports	Code execution, charts, statistics
Trader	Market analysis and trade execution	Trading engine, broker APIs, TradingView
Email	Inbox management and outreach	Gmail/Outlook API, drafting, scheduling
Social	Content creation and publishing	Platform APIs, stealth browser, scheduling
Commerce	Purchase research and execution	Virtual cards, checkout automation

Agent Lifecycle

Created → Thinking → Acting → [Waiting for Approval] → Complete/Error
              ↑         │
              └─────────┘ (loop until task done)

States

thinking — Agent is reasoning about the next action
acting — Agent is executing a tool (browsing, coding, etc.)
paused — Human paused the agent
awaiting_approval — Agent needs human confirmation for a risky action
complete — Task finished successfully
error — Agent encountered an unrecoverable error
killed — Admin force-stopped the agent

Control

Pause & Resume

Pause an agent at any point. It freezes mid-step and resumes exactly where it left off.

POST /agents/{agent_id}/pause
POST /agents/{agent_id}/resume

Guide

Inject instructions into a running agent without stopping it:

POST /agents/{agent_id}/guide
{"instructions": "Focus on the pricing page, ignore the blog"}

Cancel

Stop an agent permanently:

POST /agents/{agent_id}/cancel

Approval Gating

Risky actions require human approval before execution. The system classifies actions into 4 risk levels:

Risk	Examples	Default Behavior
Low	Read webpage, take screenshot	Auto-approve
Medium	Fill form, write file	Depends on autonomy dial
High	Send email, make purchase, execute trade	Always ask
Critical	Delete data, deploy to production	Always ask + confirm

13 action gates across 5 categories ensure agents never take dangerous actions without consent.

Agent-to-Agent Communication

Agents within a task can communicate via a typed message bus:

data_share — Share extracted data between agents
request — Ask another agent for help
response — Reply to a request
handoff — Transfer work to another agent
coordination — Sync on shared state

Memory

Each agent has access to three memory layers:

Working — Current task context (what it’s doing now)
Session — Workspace context (persists across tasks in the same workspace)
Persistent — User preferences and learned patterns (permanent)

Desktop Agents

In the Desktop app, agents gain additional capabilities:

Computer Use — Control mouse, keyboard, take full-screen screenshots
Browser Operator — Embedded Chromium with anti-detection and human-like input
IDE Operator — Open VS Code, create projects, run dev servers
Trading Operator — Open TradingView, identify setups, execute trades

​Agents

​Agent Types

​Agent Lifecycle

​States

​Control

​Pause & Resume

​Guide

​Cancel

​Approval Gating

​Agent-to-Agent Communication

​Memory

​Desktop Agents