Skip to main content
Army of Agents is a control plane for human and AI work. The web app owns company state, governance, memory, tasks, and audit history. Agent runtimes run through adapters and report back through the API.

System shape

Layers

Repository map

Request and heartbeat flow

Adapter model

Adapters keep the control plane independent from any one agent runtime. A built-in adapter can launch a local CLI, call a process, or send work to an HTTP service. Built-in adapter families include Claude Local, Codex Local, Cursor, OpenCode, OpenClaw, Gemini Local, Hermes Local, Process, and HTTP. Each adapter normally contributes:
  • server execution logic
  • configuration fields for the UI
  • diagnostics for setup and health checks
  • output parsing for run views
  • CLI formatting for terminal workflows

Data and naming boundaries

Army of Agents keeps shipped wire contracts stable. The public UI says Home, Task, Budget, Team, and Discussion. Some API routes and database tables keep older names such as /dashboard, /issues, or issues for compatibility. Do not infer that an API route name is the preferred product word. Public operator docs should use the UI language.

Governance boundaries

Key invariants:
  • every domain entity is company-scoped
  • task execution follows the single-assignee checkout model
  • governed actions can require approval
  • budget hard stops can pause agents
  • mutating actions are logged
  • memory visibility is actor-aware and scope-aware
  • agents can suggest durable memory, but approval rules decide what becomes trusted context

Where to go next

Core concepts

Learn the product vocabulary used by the architecture.

Adapter overview

Choose how your agents will run.