Unfinished Machines
Book a pilot
Pilots openSelf-improving harnesses for AI agents

your agent shipped.
we keep finishing it.

We watch your agent work, find what keeps failing, and ship tested upgrades to its harness.

Book a pilot See how it works

the harness is everything around your model.
it’s the part we improve.

Works with agents built on

Why the harness

Your agent is more than a model. It runs inside a harness of prompts, tools, examples and checks. That’s where we work: we find what keeps failing, fix it, and prove each fix before you see it.

How it works

watch. group. rewrite.
test. ship. repeat.

01Watch

it watches every run
and checks what really happened

Traces from your agent. Outcomes from your systems.

02Group

finds what keeps failing
and names the cause

Every failed task joins a failure mode.

03Rewrite

rewrites the harness
never your model

Prompts, tool descriptions, examples, validators, retries.

04Test

tests every change
against real runs

In a sandbox. If anything breaks, it never reaches you.

05Ship

ships when you say yes
and rolls back on its own

5% of sessions, then 25%, then everyone.

Your call

every upgrade arrives
ready for your yes

Approve it here, or take it as a file.

A real fix from our demo refund agent.

Results

planted defects
found and fixed

On our demo refund agent, with defects we planted on purpose.

Unneeded approvals

250

Valid refunds issued

3350of 50

Bad changes let through

0of 18

Security

your data
handled narrowly

Redacted first

Personal details removed before any model sees them.

Your own key

Raw data encrypted with a key used only for your company.

One region

Stored on AWS in Ohio. Never used for training.

Deleted on request

Erased everywhere within 30 days.

Scoped keys

Your agent’s key sends data. It can’t read results.

Instant pin

Every session back on one version in seconds.

connect one agent.
get one tested fix.

Book a pilot

Python SDK or a JSONL import. Your model stays as it is.