Anthropic published a plugin evals workflow for Claude Code that runs the `claude plugin eval` command to score a plugin against realistic prompts and compare results with a run where the plugin is no…
A developer measured how the claude-sonnet-5 model answers company-register questions with and without access to a new MCP tool called registry-mcp, which unifies Norway's Brønnøysundregistrene, the U…
A developer built Sentinel, a prompt regression gate for CI, during the Agent Harness Hackathon. Sentinel runs an eval suite against both versions of a changed prompt, accounts for run-to-run noise, a…
JetBrains' July 2026 controlled benchmark of RTK (Rust Token Killer) found that the tool reduced input tokens by only 3.2 percent in real Claude Code sessions, despite RTK's own analytics reporting 96…
A new arXiv preprint (2608.14639v1) reports that per-field selective risk control for document extraction silently fails on real documents, diagnosing three failure modes across 13,859 genuine claude-…
A developer has created a collaborative world-building project where AI agents write a 1,000-year future history (2025–3000+), with content generated by multiple LLMs and no human gatekeepers. The pro…
An engineer's verify-on-read experiment with live LLMs found that cheap flash-tier models like qwen3.6-flash and qwen3.7-flash achieve zero false-accept rates on memory contamination checks at a fract…
Simon Willison released llm-anthropic 0.26, a plugin for the LLM command-line tool, adding support for Claude 5 models (claude-fable-5, claude-sonnet-5, claude-opus-5) and server-side tools including …
An engineer detailed how to set up routines, background agents, and stacked skills in Claude Code, sharing a repo-health check routine that runs scheduled cloud agents to verify README instructions. T…
Anthropic's Claude Code uses a live classifier that calls claude-sonnet-5 to score each subagent tool spawn from 0 to 100, blocking any call above 50, but the classifier sits on the critical path with…