Three studies examine analytical choices, search coverage, and structured retrieval. We relate them to two Foam evaluations of tool calls and incident diagnosis. None measures a general change in answer time.
We measured whether individual tool calls helped coding agents solve observability tasks. Across two models, removing discussion and issue tools cut calls by 58%; task accuracy held steady or rose.
Across 63 tasks in 25 evaluation runs, Foam’s filesystem-based Recursive Language Models cut token spend by roughly 90% with no observed loss of accuracy. The evaluation changed storage, delegation, and model choice together.
On 22 production incidents, Cursor diagnosed 12 root causes alone, 14 with the Foam tool, and 9 with the Sentry tool. A separate Foam agent using the Pi harness diagnosed 18; its prompts and harness also differed from Cursor’s.
On 22 production incidents, Claude Code scored 19 correct diagnoses with Foam, 14 with repository tools alone, and 12 with Sentry. Each setup ran once per incident.
In migration data from 11 production environments, Foam produced 95.5% fewer issue records than Sentry. The median environment had a 10.9× reduction; clustering quality was not measured.