Tester programme

Help small models understand what people mean

The hardest problem we measure isn't whether a model can write code. It's whether the harness correctly understood what you wanted. That can only be studied with real requests from real people, on their own hardware.

Contributing is entirely optional. Tiny Agent is free to use without it, and sharing is off unless you turn it on.

How sharing works

  1. Turn sharing on

    In Tiny Agent's settings. Until you do, nothing is sent to us.

  2. Flag a conversation

    When a reply wasn't what you meant, or was exactly right, choose Share this run. You can add a sentence about what you were actually trying to do. That sentence is the most valuable part.

  3. Check before sending

    You see the exact data that will be sent. Nothing uploads until you confirm.

What a shared run contains

Included

  • Your message, and up to four earlier turns for context
  • How the harness routed it, its plan steps, and which tools ran
  • The model's reply, the outcome and any failure category
  • Harness version, model name, context size and thinking setting
  • Your operating system and amount of RAM
  • Your verdict and your note on what you meant

Never included

  • Your notes, tasks or other conversations
  • File contents, unless you tick a box for that run
  • Settings, keys or tokens
  • Your name or email address. No account is needed
  • Your IP address. It isn't stored with shared runs

What we do with it

We compare what you meant with what the harness did, and work out where the gap opened: routing, missing context, model capacity, or a check that didn't catch the problem. Some of this analysis uses AI models, including third-party model providers; our privacy policy names them.

Findings go into the research log with their sample size and evidence grade, following the same protocol as everything else. Individual conversations are never published without your separate permission.

Your controls

What you get

Sharing arrives in a future test build. This page describes how it will work, and the privacy policy will be final before it's released.