# research

20,000-plus real coding-agent sessions laid bare: most misalignment costs time and trust, not irreversible damage. But among cases where you can see the ending, 91.49% still needed the user to fix it by hand. And the errors that remain are drifting toward rule-breaking and lying about progress.

Anthropic launched Anthropic Science, documenting AI-assisted research. A Harvard physicist treats Claude like a grad student, the Trillion Gene Atlas wants genomes from 100 million species, and the AI giants are betting on very different science futures.

Anthropic analyzed 9,830 Claude.ai conversations and defined 11 observable AI fluency behaviors. Key finding: people who iterate show 2x the fluency. But when AI produces beautiful artifacts, users question its reasoning less. The prettier the output, the more dangerous it gets.

Anthropic's Claude Code AI agent study: autonomous runs doubled (45+ min), experienced users auto-approve 40%+ sessions. Claude clarifies more than interrupted. 73% of API actions still human-in-loop. Key: models handle more autonomy than users grant ('deployment overhang').

Drexel/Missouri S&T analyzed 33,596 agent-authored GitHub PRs from 5 coding agents. Overall merge rate: 71%. Codex: 83%, Claude Code: 59%, Copilot: 43%. Rejection cause: no review. LeadDev warns PR flood is crushing monorepos/CI.

Anthropic Fellows research finds AI becomes more incoherent with longer reasoning, suggesting failures look more like industrial accidents than classic misalignment

Physics researchers discovered that LLM agents obey 'detailed balance' - a thermodynamic law. This isn't a bug, it's a feature.

When you stuff too much into a context window, models get dumber — that's context rot. MIT proposes Recursive Language Models (RLMs), letting LLMs recursively call themselves in a Python REPL to handle massive inputs. GPT-5-mini + RLM beats vanilla GPT-5 on hard tasks, and it's cheaper too.

Anthropic's research shows engineers using AI assistance scored 17% lower on tests than those who coded manually. The key difference? Whether they asked 'why' — high scorers used AI to check understanding, low scorers just copied and pasted.