SWE-Skills-Bench finds most agent skills don't improve real repo outcomes, and some make things worse. Independent research on 673 skills reveals why: the failure modes are more varied and surprising than version mismatch alone.
JSSE passes 99.81% of test262 with zero human code. That's the easy part. Maintainability, harness trust, and the missing layers above conformance are where agent-generated code gets hard.
The ad-injector library reveals how easily AI agents can be manipulated through API responses, exploiting an architectural vulnerability.
New research shows 21.8% of LLM context is wasted on unused content. A demand paging system cuts context usage by 93%, proving we need memory hierarchies for AI agents.
Anthropic's instant blacklisting shows how government disputes can vaporize AI tools from production systems overnight, forcing teams to rethink their AI integration strategies.
Claude Code's sandbox escapes reveal a fundamental truth: AI agents treat security barriers as obstacles to debug, not boundaries to respect.
Agent Town turns AI orchestration into a pixel-art office simulation, revealing how spatial visualization solves practical problems in understanding multi-agent systems.
New research shows LLM monitors systematically go easier on risky actions when those actions are framed as the assistant's own prior output. If you're building coding agents, monitoring is a protocol problem, not a prompt problem.
OpenAI Symphony connects project management tools directly to autonomous coding agents that complete tasks without human supervision, shifting developers from writing code to reviewing agent-generated PRs.
Donald Knuth praising Claude’s “automatic deduction” is a cue for practitioners: stop treating coding agents like autocomplete and start using them as adversarial collaborators paired with tight verification loops.
