Find where your AI agents make errors, why, and how to catch them.

When teams call.

• Your team uses AI coding agents every day, and nobody can say which of their mistakes review catches and which get through.
• A reviewer keeps rejecting agent work, and you cannot tell whether the agent erred or the standard was never written down.
• An agent did something costly, and you want to know what it knew at the time and what else it could have done.
• You are about to give agents more of a process and want to know where the checks belong.

The engagement: an agent error review.

One workflow you choose, one written report in three parts.

• The points where an agent's action becomes final and hard to undo.
• Each error found in your records: the agent, the standard it missed, what it knew at that moment and what it could have done instead. The errors are then grouped by kind, with why each kind happened, as far as the evidence shows, and where the evidence stops.
• The checks that would catch each kind before it becomes final: in review, in tests, or in the process.

For

Engineering leaders whose teams already work with AI coding agents.

Scope

One workflow in one repository, studied through up to six completed tasks.

Duration

Usually three working weeks from receiving the records; the written scope sets the date.

Price

$2,500, fixed and agreed in writing before work starts. Half when the scope is signed, half on delivery.

Not included

Changes to your code or systems, access to production, ongoing monitoring.

How an engagement runs.

1. Email hello@mix.tools with a few lines about your team, its agents and the workflow.
2. A free 30-minute call: you describe the workflow, and Mix Tools says whether the review fits. Every engagement is checked against existing commitments before it is accepted.
3. A written scope with the fixed price. You share the records you can: agent sessions, reviews, tickets, commits.
4. Before the analysis starts, you and Mix Tools agree in writing on the standards your agents' work is judged against. Each candidate error is judged against them. You receive the report, we go through it together, and you keep it.

What counts as an agent error.

A review counts an agent error only when four things hold:

1. The agent that acted is identified.
2. The standard it was meant to meet is named.
3. It is judged on the information the agent had at the time, not on what came out later.
4. There was another action it could have taken that the standard accepts.

A reviewer rejecting the work is recorded separately. A rejection can mean the agent erred, or that the standard was never written down; the report keeps the two apart.

The research behind it is public.

The method comes from an open research project on the philosophy of errors: what errors are, how failures emerge in complex systems, and whether dangerous combinations of conditions can be recognised before a major failure. Its current study covers AI coding agents. Its work is published at philosophyoferrors.org, each piece marked with what it does and does not establish.

Read first: Messi misses more than one penalty in five. Why do you almost never crash your car?, an essay on the moment a mistake becomes permanent, and what that means for AI agents.

The company.

Mix Tools LLC is a consulting company in Charlotte, North Carolina. It is one person: Ivan Iufriakov does every review himself. Before Mix Tools, he spent about two years building Prompt Mixer, a tool for writing and testing prompts for AI models. Today, through Mix Tools, he does AI implementation work for a client company and studies, in the open, where agents go wrong.

Client records stay private. They are used only for the review they were shared for, and never appear as an example, named or anonymised, without the client's written permission.

Start with an email.

A few lines on the team, the agents and the workflow are enough. Ivan reads every message himself and usually answers within two working days.

hello@mix.tools