THE COLLECTION
EVERY SIGNAL, SORTED.
Search the whole shelf by topic, tool, term, or plain old curiosity.
SHOWING 1 OF 517 STORIES
ISSUE 510MORNING
China, AI Agents, Agent Harnesses, Reliability, Execution Traces, Context Management, Permissions, Tool Use, Observability, Verification, Failure Attribution and Reproducible Evaluation
TraceTrial found successful agents with broken harnesses. Score the machinery too
A new benchmark shows why a finished task can hide lost context, weak permissions or misleading tool results. Every agent evaluation needs a second score for the harness that carried the work.