The science of agents: planning, tool use, memory, and evaluation. From ReAct and Reflexion to the benchmarks that measure whether an agent finishes the job, and what each run costs.
When a long run goes wrong, most teams can kill it or argue with it. Both are bad. A steer is a third option: a message with a delivery mode, a status, and a record.
METR measures how long a task an agent can finish on its own. That number has doubled roughly every seven months since 2019. What a week-long run does to supervision.
The worry about AI slop is about output quality. The measurements point somewhere else: at teams handing an agent a human workflow and expecting it to hold.
A prompt is a bad container for a standing rule. It has no owner, no date, no scope, and no record that the agent read it. Context records fix the container.
Self-correction fails without external feedback, and training on your own output degrades the model. What the collapse and verifier research says you should do instead
A coding agent changed your tree and nobody can replay how. What the research says about execution feedback, sampling variance, and turns you can audit
