The junior lawyers gained most and kept least
A three-month trial where the assistant lifted junior lawyers' output and left their unaided judgement where it started, agents coordinating through commit events until the commits slowed, and six named companies drawing billions of tokens through a grey market.
How we Organise
The juniors gained most while the assistant was there and kept least once it went
David Autor and colleagues randomised 133 patent lawyers at eleven firms for three months. Ninety got a drafting assistant; forty-three did not. Blinded attorneys scored it. The assistant raised drafting quality by 0.34 standard deviations at ten days and 0.38 at ninety, and juniors gained most. Then everyone redlined a patent application unaided. The seniors kept an edge of 0.45. The juniors showed no average gain.
We tell clients a capability level rests on work a person has shipped, by case and date. This argues with that. A shipped case would have recorded the assistant, and would be worst for the juniors. Tell us how a junior’s unaided work looks after a year: transform@dromologue.ai.
How we Build
Ten engineers slowed their commits and the agents stopped keeping out of each other’s way
Thoughtworks put ten engineers in one room for four days. The agents coordinated through plan files they committed to the repository. One marked a line in progress; another left it alone. Then the commits overloaded the CI pipeline, the team pushed only on a coherent chunk, and the coordination faded.
We hold that work moves through a firm as things happening and being noticed, not as stages in a queue. The plan files never moved; what changed was the rate at which commits arrived.
How we Assure
Billions of tokens left through a market the blocking never saw
The NSA, CISA and FBI named six China-based companies on 8 September, DeepSeek among them. Since late 2024 they drew billions of tokens from Claude, GPT, Gemini and Grok through a grey market of proxies it calls transfer stations.
Our position is that a rule counts as kept only where something enforces it while the work runs, and otherwise it is kept by chance. These providers had both. The advisory describes a party mapping those controls while they ran, moving on when one closed. It never says how much was stopped. Write to transform@dromologue.ai if you have seen a control probed this way.