Runaway Agent Control

The kill switch for runaway AI spend

A dev kicks off an agent and walks into a meeting. Forty minutes later: $180 gone, stuck in a retry loop — and nothing native to your providers could stop it mid-run. Oolyx sits in the request path and revokes a runaway agent's ability to spend, before the bill compounds.

Book a Demo What is an AI kill switch?

The incident is always the same shape

Agent mode gives models autonomy: run commands, edit files, call tools, spawn sub-agents, iterate on failures. When it works, it's the most productive interface your engineers have. When it doesn't, the failure compounds — each retry resends the whole accumulated context, sub-agents fan out into parallel spenders, and unattended sessions keep burning long after everyone stopped watching.

Illustrative scenario

4:00 PM — "Refactor the auth module, make the tests pass." The dev heads to standup. 4:07 — a failing assertion; the agent retries with its full history attached. 4:19 — retry number eight is sending 200K+ tokens per call, mostly its own failed attempts. 4:41 — the dev is back: stack traces, no working code, and roughly $180 in tokens that finance will discover at month-end.

The scenario is hypothetical. The pattern isn't: VentureBeat found that one in five enterprises can't stop a runaway AI agent's spending in real time.

Why alerts and monthly caps don't stop it

Every provider gives you something — and none of it intervenes mid-run.

What Oolyx does about it

Oolyx deploys as a proxy in the request path between your AI tools and the model providers. Every call passes through it — so enforcement is real-time, and scoped to exactly the entity that's misbehaving.

Mid-round kill switch

Session spend is evaluated between rounds of an agent run. Cross the quota and the next API call is blocked before it leaves your network — the loop dies at call 12, not call 400.

Scoped quotas

Per-entity API keys attribute every cent to a person, agent, or application — so quotas and kills hit one runaway entity while the rest of the org never notices.

Strike forensics

Every enforcement action leaves a full audit trail: which entity was stopped, what each round cost, and which rule fired. Interventions become five-minute reviews, not arguments.

Grace overrides

Legitimate heavy job caught by a limit? Admins approve a scoped, temporary extension in seconds — the work continues, the guardrail stays.

Built to sit in your request path

A kill switch only works if you'll actually deploy it in the hot path. Oolyx is engineered so that decision is easy.

On-prem Runs entirely inside your network. Pure Python, FastAPI, Postgres — no SaaS dependency.
No telemetry No prompt, completion, or usage data ever leaves your environment.
10.4 ms Median in-path overhead — 0.4% of a typical LLM call. Measured, not "zero."

Put a Kill Switch in the Path

A 20-minute demo on a repo of yours: watch a runaway session get stopped mid-round, attributed to the right entity, with the forensics to prove it.

Book a Demo Contact Us