The Agent Field Guide
Catalog

Researches

One folder per research, one standalone page per report, with the brief and the evidence trail next to it. Newest first.

1 research

  1. Thirteen days with an agent

    What 13 days of working with the OMP coding agent actually looked like, measured from the local session store rather than from memory. 285 human messages produced 5,295 main-agent turns, and 17,847 across every role; a background advisor that never called a tool took 67.1% of the bill; and "how long did it run" has five different correct answers. Ends with the model, capability and workflow changes the numbers argue for.

    · agents · telemetry · cost · self-measurement · recommendations · complete

  2. Memory systems for LLM agents

    Which systems let an agent retain facts, connect them, reason about temporal data, and apply memory decay. 25 systems scored against six requirements, 100 sources, interactive decay models.

    · agents · memory · benchmarks · architecture · complete