An archive of decisions, failures, changes, and shipped work. The record stays useful by keeping the sequence—and the limits—visible.
How harness, tool, attempt, and judging conditions become part of what an agent benchmark score actually describes.
FineEnvs reports the same LFM2.5-2.6B model weights scoring 62% with Mini-SWE-Agent and 33% with Claude Code on the held-out tasks in its multi-harness experiment .
Why AI operating cost is easier to reason about when context, routing, model choice, deployment placement, and measurement are treated as separate controls.
A smaller agent tool catalog can simplify invocation while relocating the real execution contract into authorization, containment, observability, recovery, postcondition evidence, and acceptance authority.
How sandbox lifecycles turn agent deployment boundaries into recovery decisions about state survival, restartability, provisioning latency, and replay.
Repeated agent turns can turn reusable context into a latency and capacity problem when cache locality, prefill pressure, and live load stop lining up.
Why testing MCP tools requires a separate semantic behavior check after protocol and schema validity, so accepted requests are tested for the effects their parameters promise.
Long-running agent recovery depends on preserving execution state, pending decisions, artifacts, and side-effect history rather than simply retrying the last request.