Metatransformer · Public notebook
Short working ideas from active book research: published early, linked to their sources, and allowed to change as evidence and counterexamples sharpen them.
Follow the RSS feedPublished notes · 6
Corrections and useful responses move forward into the editorial process; the original note remains a working artifact, not a finished chapter.
5 min read · Research note
Long-running work changes which tasks are ready while turning completed behavior into regression obligations that every later edit must preserve.
Topics: agents, software engineering · Tags: long horizon agents, dependency dags, regression evaluation
5 min read · Research note
An executor may propose that work is complete, but only independent observation of the environment should advance durable task state.
Topics: agents, distributed systems · Tags: task state, environment verification, agent recovery
5 min read · Research note
Pass rate belongs to a model, harness, environment, and budget configuration; the model name alone hides cost, latency, failure shape, and oversight.
Topics: agents, ai engineering · Tags: agent harnesses, evaluation cost, system procurement
4 min read · Refinement note
Runtime choices matter, but the useful question is which contracts keep a trajectory controlled when proposals, effects, evidence, and the system itself can change.
Topics: agents, ai engineering · Tags: agent harnesses, system contracts, book thesis
4 min read · Research note
After a timeout, replaying the conversation cannot tell you whether an external effect happened, which operation it belonged to, or how to reconcile it safely.
Topics: agents, distributed systems · Tags: durable state, recovery, external effects
3 min read · Research note
A convincing patch description is still a proposal until the environment changes and independent observations establish the result.
Topics: agents, software engineering · Tags: outcome evidence, repository state, agent runners