title: Introduction order: 1

Introduction

AAS (Autonomous Agent System) is a self-improving agent swarm that detects infrastructure problems, fixes them automatically, remembers the solutions, and improves over time.

What It Does

After startup, AAS agents run in the background without human intervention. They:

  1. Detect issues — uncommitted changes, high error rates, services down, log errors
  2. Check cache — have we solved this before?
  3. Analyze & plan — if new, use LLM to understand and decide
  4. Execute — run actual commands (git commits, log cleanup, service restarts)
  5. Verify — confirm the problem is fixed
  6. Learn — cache the successful solution
  7. Improve — adjust confidence thresholds based on success rate

Real Autonomy

Unlike simulation-based tools, AAS runs actual commands and verifies they work:

  • Git: Commits, branch cleanup, pushes
  • Logs: Rotation, error analysis, cleanup
  • Services: Health checks, restarts, alerts
  • Metrics: Memory/disk cleanup, optimization

Failures are rolled back. Successes are cached and reused.

Cost Model

  • First occurrence: 2 LLM calls (analyze + plan)
  • Repeat occurrences: 0 LLM calls (cached)
  • Example: 100 cycles, 10 unique issues → ~90% LLM savings

This is where AAS saves money. Learn the solution once, reuse it forever.

Multi-Provider LLM

AAS intelligently routes tasks to different LLM providers:

  • Claude API — deep reasoning (analyze, plan)
  • Hermes — fast local analysis
  • Claude Code — file edits and code modifications
  • OpenClaw — external integrations
  • Fallback — mock responses if all else fails

Agents use the fastest, cheapest option for each task type.

When to Use AAS

  • Environments that need constant, hands-free monitoring and fixing
  • Teams who can't afford 24/7 ops staff
  • Codebases with predictable, repeatable problems
  • Infra that benefits from experience-based improvements

Next Steps