Production AI agents loop, burn tokens, and have no kill switch. RiskKernel is an open-source, self-hosted runtime that puts hard limits around any agent: per-run cost, loop, and time budgets, all enforced in Go — never by an LLM. Point an existing agent at it with one env var. Your keys, your infra, no telemetry. Plus crash-resume, human approval gates on side-effecting tools, and OpenTelemetry export. Apache-2.0. The LLM proposes; deterministic code disposes.

Maker here 👋 I built RiskKernel after years of building deterministic risk engines around non-deterministic systems — the kind where a mistake costs real money. The lesson that stuck: the thing keeping those systems safe was never the smart part, it was the hard-coded layer around it. Agents hit the same failure list everyone knows — runaway loops, surprise token bills, no kill switch, no recovery. Frameworks orchestrate the reasoning but ship none of the guardrails. RiskKernel is that guardrail layer: cost/loop/time budgets enforced in Go, one env var to adopt, your keys, no telemetry. It's early (v0.2) and honest about its limits. I'd genuinely love your feedback on where the guardrails are too strict or too loose.