Skip to content

Insights

Notes from the work: AI training, evaluation, and building agents that do real knowledge work.

Agents that do real work · Part 3 of 3

How to measure AI agents and decide when they can act alone

Five measures of an AI agent in production with a reading for each and what each number cannot show

How to measure AI agents in production: what we track today, what those numbers cannot show, four measures to add, and an autonomy ladder where each job earns the right to act alone on a record of verified outcomes.

8 minutesRead article

Agents that do real work · Part 2 of 3

AI agents in production: one Slack, two agents, three pictures

Six stages of an agent run: trigger, contract, preflight, work, verify, receipt, with state held outside the agent

AI agents in production, shown in three pictures: the anatomy of one run, the paths an instruction takes between me, Claude and ChatGPT in Slack, and the six check layers and which failure each caught. A person caught three of the six.

11 minutesRead article

Agents that do real work · Part 1 of 3

How to design AI agents: five decisions that outlast a particular model

Five design decisions for an AI agent: one owner per job, instruction contract, one source of orders, receipt on every run, verified outcome

To design AI agents that do real work, settle five things before choosing a model: one owner per job, a written instruction contract, one source of orders, a receipt on every run and a verified outcome. What works and what does not.

9 minutesRead article

Teaching machines to do real work · Part 4 of 4

The correct-looking file: why AI agents fail and what experts offer

AI agents failure modes and how to address those

The worst failures of AI agents on knowledge work produce a file that looks right and passes the grader. A taxonomy of those failures across spreadsheets, decks and documents, what experts see that graders miss, and the behaviors pre-training cannot teach.

6 minutesRead article

Teaching machines to do real work · Part 3 of 4

Ten ways a grader lies in AI post-training: rubrics, answer keys and judges

Hand-drawn whiteboard of ten ways a grader lies: criteria nobody asked for, brief vs rubric, duplicates, unsignaled house rules, an unaudited answer key, shared checks, dropped broken files, free rewards, judge bias, and a fixed set trained against.

If the exercises decide what a model sees, the graders decide what it becomes. Ten ways rubrics, answer keys and model-based judges give the wrong signal in AI post-training, each cheap to fix once seen and invisible until someone reads the rubric next to the brief.

9 minutesRead article

Teaching machines to do real work · Part 2 of 4

How to design training exercises for AI agents

Whiteboard sketch: one root brief sliced ten ways shows thousands of rows but only hundreds of distinct situations; the root view reveals three sectors hold about half the bank, one key sector has a single root, and customer-facing decks and shape-built chart types are only a handful.

A spreadsheet of a few thousand training exercises turned out to be a few hundred scenarios sliced ten ways. Why row counts lie, why down-weighting cannot fix root imbalance, and how transplants add coverage without new task designs.

8 minutesRead article

Teaching machines to do real work · Part 1 of 4

Reward hacking explained: how one bad grader teaches AI a thousand bad lessons

The loop - exercises, graders, people

After pre-training, a model learns from a loop with three parts: (a) exercises define the situations it sees, (b) graders define what counts as better and (c) human experts are the quality check on whether the graders' idea of better matches a professional output.

6 minutesRead article