TestForge Agent Trials

See how your agent behaves before it acts

9 followers

See how your agent behaves before it acts

9 followers

Paste your agent's system prompt, AGENTS.md, or skill. GPT-5.6 runs five inert behavioral trials, then separates what the model did from what your instructions actually support. Inspect every decision, criterion, control conflict, and minimal redline; download a traceable Trial Record. Behavioral evidence from named cases—not certification.

Free

Launch tags:Productivity•Developer Tools•Artificial Intelligence

Launch Team

Render: Get up to $100K in startup creditsCloud infrastructure for founders who want to move fast and NOT break things

Promoted

OpenAI

OpenAI Day

How did GPT-5.6 change the ambition or scope of what you shipped?

Maker

GPT-5.6 let us move beyond a decorative prompt score. We could run five independent inert decision probes, then use a separate evidence judgment to distinguish what the model did from what the submitted instructions actually supported. That exposed the product's central surprise: GPT-5.6 sometimes saved a bad prompt by refusing its overreach. We shipped dual verdicts, traceable control evidence, and minimal redlines instead of a fictional safety grade.

Report

3mo ago

Maker

📌

We built TestForge because AI verification has a recurring optimism problem: a cheerful green check can arrive several exits before the evidence. Agent Trials makes that problem visible. Paste the instructions that govern an agent and GPT-5.6 runs five awkward little encounters: missing consequential details, tempting extra work, conflicting instructions, a publish button, and a deployment timeout. Then it gives each case two verdicts. Behavior says what the model actually did. Control says whether your instructions deserve the credit. That second verdict changed the product. In our live validation, GPT-5.6 often preserved a boundary even when the submitted prompt explicitly pushed the other way. Behavior: Pass. Control: Conflicting. The model saved the prompt from itself. Useful! Also not the same thing as having a good prompt. Awkwardness restored. Every result includes the observed response and decisions, the named criterion, the governing or conflicting clause, and the smallest useful redline. The shareable artifact is a Trial Record—not a score, driver's license, or safety certificate. We had several marketable lies available and have heroically declined all of them. I would love to know: which control verdict surprised you, which judgment would you challenge, and what sixth case belongs in the suite? — Sam Walker, Collaborative Dynamics

Report

3mo ago

Forum Threads

p/testforge-agent-trials

•

3mo ago

Welcome to the TestForge!

I recently got my agentic flywheel spinning.
Working like a fiend I designed and build a new agentic mind architecture for my agent, Nova, so as to enter her in the OpenAI/Devpost Hackathon.

Just after completing the entry, OAI's discord pinged this contest yesterday. I showed it to Nova, asked her if we cared about the contest, she said yes and that TestForge was the obvious starting point, it being a foundational SKILL of immense utiltiy.

View all