Skip to content

AGENTS.md

Latest commit

 

History

History
51 lines (38 loc) · 1.96 KB

AGENTS.md

File metadata and controls

51 lines (38 loc) · 1.96 KB
Raw

AGENTS.md — Agent-OS Elixir Agent Loop

Agent Loop

Every AI coding agent working on this codebase follows this loop:

Dev → Compile → Fix → Write Tests → Test → Notify

Steps

  1. Dev — Read the task, explore the codebase, write or modify Elixir modules
  2. Compile — Run mix compile --warnings-as-errors from the app directory
  3. Fix — If compilation fails, fix all errors and warnings, then go to step 2
  4. Write Tests — Write ExUnit tests for the changes (real OTP processes, no mocks)
  5. Test — Run mix test from the app directory
  6. Notify — Report results: which files changed, tests passed/failed, warnings

Rules

  • All tests use real OTP processes — no mocks, no stubs, no fakes
  • Every GenServer under test must be started and stopped cleanly
  • Tests must be deterministic (no :rand without seed, no timing-dependent assertions)
  • Compile with --warnings-as-errors — zero warnings allowed
  • Each umbrella app is compiled and tested independently
  • Integration tests live in src/agent_os/test/

Umbrella Structure

src/
├── agent_os/          # Main orchestrator (depends on all subsystems)
├── agent_scheduler/   # Agent lifecycle, scheduling, pipelines
├── tool_interface/    # 3-tier tool registry, capabilities, sandbox
├── memory_layer/      # ETS + Mnesia typed memory
├── planner_engine/    # Order book, escrow, reputation, market
├── agent_os_web/      # REST API (Plug + Cowboy)
└── agent_os_cli/      # CLI (escript)

Compilation Order

memory_layer → tool_interface → agent_scheduler → planner_engine → agent_os → agent_os_web → agent_os_cli

Agent Types

  • OpenClaw — Full-capability research agent (web_search, browser, filesystem, shell). Default oversight: :autonomous_escalation
  • NemoClaw — NVIDIA-secured agent with restricted tools, privacy routing, policy guardrails. Default oversight: :supervised