Agent frameworks are getting very good at doing things — and much less good at not doing things. That gap is where the damage lives. It's why we initiated an open-source project to close it — and why we're contributing it to the communities that need it rather than keeping it in-house.
Open Octi is an agent harness that sits between an AI agent and the systems it can touch. Every plan an agent produces surfaces to a gate before anything executes. No approval, no execution. A cancelled plan executes zero side effects — not "probably zero," zero. Fail-closed is the design contract, not a feature flag.
What exists today
A hardened harness core with a real completion contract. Plan gating verified live against a real gateway. A plugin that registers Open Octi as an agent harness for OpenClaw — the first of the agent communities we want this working in, not the last. Even the ops scripts follow the philosophy: the process kill-guard refuses ambiguous targets, and has already refused one in the field, exactly as designed.
The rules that never change
Fail closed — ambiguity is a denial. Live systems are read-only to anything experimental. Identity before action — processes are verified by what they own, not what they're named.
Why we're behind this
Our commercial platform puts AI staff inside real businesses — answering phones, filing records, touching money-adjacent systems. We will not ship autonomy without a gate we'd bet our own firm on, and a gate like that is worth more audited in the open than trusted in the dark. So the harness is MIT-licensed, developed publicly, and honest about its maturity: it's early, interfaces will change, and you shouldn't point it at production systems you care about yet. That honesty is the point.
Where it goes next
A planner proxy with policy gating — standing approvals by plan shape, automation envelopes, re-gating on deviation — then a sandboxed lab environment and a model-call classifier. Harnesses for more agent ecosystems as the core hardens.
The repository is the front door. Questions, skepticism, design arguments, and contributions are all welcome — issues get read and answered.