How to install

The agent stopped making progress. The bill did not.

Gubernaut spots an agent stuck in a loop and orders the stop before the next call.

On a verbatim loop the stop lands at turn 4, and turns 1 to 3 are sent and billed.

Install nowRead the research

One controller. Two ways to run it.

Gubernaut Tiller

Tiller is a local proxy that speaks the OpenAI API. It stops a looping call itself.

pip install gubernaut-sdk

Language-neutral: any client that speaks the OpenAI wire format can point at it. A proxy process must be listening first.

Tiller on GitHub (opens in a new tab)More about Tiller

Gubernaut Keel

Keel is a library inside the program. It decides, and the program does the stopping.

npm install @gubernaut/core

It decides, it does not intercept. Acting on a REGROUND posture is the caller's job, which is the one thing the proxy does that this does not.

Keel on GitHub (opens in a new tab)More about Keel

Both are free and open source, under Apache-2.0.

Install now

Verbatim loop batteryGubernaut Tiller

79.8%to95.9%of a runaway loop's bill, saved

Measured onGPT-5.6 Sol, GPT-5.6 LunaClaude Fable 5, Claude Haiku 4.5Gemma 4 26BLlama 4 Scout

Measured through the proxy, Gubernaut Tiller, with both arms making the same number of attempts. Flagship: gpt-5.6-sol, $0.1669 ungoverned against $0.0068 governed, per 25-attempt loop.

See the receipts

Pre-registered ablation, 2026-07-21, gpt-5.6-sol, N=3.

117%to192%of the unguarded bill, spent by a calming prompt

See the measured breakdown

Spend against an unguarded baseline

Spend on the same saturating loops, as a share of what they cost with nothing governing them.

Gubernaut: 23% to 63% of the ungoverned baseline

A strong calming system prompt: 117% to 192% of the ungoverned baseline

On these RLHF-aligned frontier models the governed-vs-prompt-only warmth-recovery contrast was weak and mixed. No behavioral-tone headline is claimed.

Source: Pre-registered ablation, 2026-07-21, gpt-5.6-sol, N=3.

Gubernaut

Pre-registered ablation, 2026-07-21, gpt-5.6-sol, N=3. The static prompt was insufficient on 3 of 4 frontier families. On these RLHF-aligned frontier models the governed-vs-prompt-only warmth-recovery contrast was weak and mixed. No behavioral-tone headline is claimed.

15 of 16model pairs came back calmer with the governor on

Tested across four model families, each one's work judged by all four.

13 of 16 reached p<.05. One pair showed no effect: -0.04, GPT×Gemini.

Read the researcharXiv 2607.24339 (opens in a new tab)Paper, PDF

DOI 10.5281/zenodo.21303518 · ORCID · data under CC-BY 4.0 · the evidence release

Three things it does

  • When an agent repeats the same request, the fourth try is stopped and never billed.

    Verbatim-loop battery, seven measured configurations across four model families. Turns 1 to 3 still spend. Measured through the proxy, Gubernaut Tiller.

  • If the proxy goes down, the agent stops. It never carries on unguarded.

    Measured against a canary upstream in the 0.1.1 hardening pass and re-run at HARDTEST v1.0. The battery is the proxy's and its clients'.

  • Free and open source. It runs where it is installed and reports nothing home.

Free, open source and self-hosted.

Apache-2.0. Questions go to the engineer who built it.

Install nowWrite to the lab