Myles — the circuit breaker for AI agents

Stop your agent bankrupting you overnight. A zero-config local proxy that catches runaway loops and budget blowouts before they burn money — any agent, any language, any provider, no code changes.

pip install myles-ai
myles demo            # see it work in 15s (no API key)
myles run             # start watching your agents (observe-only, safe)

The problem

Autonomous agents get stuck — a broken tool retried thousands of times, a reflection loop that never converges, dozens of parallel sessions grinding the same impossible lookup. Developers wake up to $2k-in-4-hours and $8k-overnight bills. Platform daily caps kick in after the runway is gone. (real stories.)

Myles sits in the request path and stops the bleed in real time.

See the value in 15 seconds

myles demo

  call  5  forwarded  $0.0153   <-- MYLES: runaway loop detected!
  call  6  BLOCKED    $0.0000   (Myles refused — would have cost $0.0153)
  ...
  SAVED : $0.1530 on this ONE loop

Try it on your own agent (safe — observe-only)

myles run                                   # observe mode: logs, never blocks
export ANTHROPIC_BASE_URL=http://localhost:8788   # Claude Code
# ...use your agent normally...
myles report                                # what it WOULD have stopped + saved

When you trust it, flip on enforcement:

myles run --enforce --budget 5.00           # now it actually blocks

What it stops

Not a compressor. Not a router. Not a dashboard. A wire-level kill switch that refuses the bad call mid-flight and shows you what it saved.

Why it works (30s)

The detector measures information novelty per step: a healthy loop keeps exploring new state; a runaway collapses to near-zero novelty while still spending. Similarity is IDF-weighted (learns your boilerplate, so read file src is ignored and the real intent decides) and response-aware (only a proxy sees responses — that's what separates a refining loop from a stuck one). Built and validated with an open probe — see probe/FINDINGS.md.

Install

pip install myles-ai   # zero dependencies, stdlib only, Python 3.9+ · provides the `myles` command

Or run from source: python -m myles run.

Status

v0.1 — early MVP, seeking feedback. Enforcement validated end-to-end. Honest limits:

• Observe-only by default on purpose — try it with zero risk, then --enforce.
• Cross-run uses a coarse intent key; can be sensitive on common short intents (tuning next).
• Streaming is buffered (fine for coding-agent sizes); true pass-through is a later optimization.

Issues and PRs welcome. Apache-2.0.

Let's connect

I'm building Myles in the open and I'd genuinely love to hear from you — whether it's feedback on the idea, a bug, a horror story of your own agent going rogue, or you just want to talk about the project. Reach out:

• LinkedIn: https://www.linkedin.com/in/elam-olame-mugabo/
• Website: https://elamolamemugabo.com/

— Elam Olame Mugabo

Built for people who ship agents and hate surprise bills.