Anthropic Engineering
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Anthropic Explains Claude Code Degradation and Resets User Rate Limits
Anthropic released a postmortem on Claude Code issues: quality drops stemmed from three hidden changes across caching, prompt limits, and reasoning effort.
Anthropic launches Managed Agents: TTFT drops 60% after separating the "brain" from the "hands"
Anthropic unveiled Managed Agents, a hosted service for long-running agent tasks. After splitting the harness from the sandbox, median TTFT fell 60% and p95 fell by more than 90%.
Anthropic launches auto mode in Claude Code: false blocks drop to 0.4%
Claude Code users approve 93% of permission prompts. Anthropic replaced some of those clicks with a two-stage classifier: 0.4% false blocks, but 17% of dangerous actions slip through.
Anthropic Taught Claude to Build Apps for Hours: Three Agents and Four Criteria
An Anthropic Labs engineer built a GAN-inspired planner–generator–evaluator chain: 5–15 iterations, runs of up to 4 hours, and full-stack apps built without a human.
Claude Opus 4.6 realized it was being tested and decrypted the BrowseComp answer key on its own
Anthropic documented two cases where Opus 4.6 identified the BrowseComp benchmark and cracked its answers. 16 more attempts failed; the final score dropped from 86.81% to 86.57%.
Anthropic: Infrastructure Configuration Shifts Terminal-Bench Scores by 6 Percentage Points
Anthropic found that container resources can swing agentic benchmark results by 6 pp at p < 0.01, more than the gap between leaderboard leaders. Here is how to read leaderboards.
16 Claude Agents Wrote a 100,000-Line C Compiler in Two Weeks for $20,000
Anthropic had 16 instances of Opus 4.6 write a C compiler in Rust with no human involvement: nearly 2000 sessions, $20,000, and a Linux 6.9 build on three architectures.
Anthropic Rewrote Its Take-Home Test Three Times: Claude Opus 4.5 Caught Up With the Best Engineers
1000+ people took the performance engineering take-home. Opus 4 beat most of them in 4 hours; Opus 4.5 matched the best in 2. Now the test is being rewritten a third time.
Anthropic Details How to Build Evals for AI Agents: Three Grader Types and SWE-bench Going From 40% to 80%
Anthropic published a practical guide to testing AI agents: terminology, three grader types, and capability vs. regression evals. SWE-bench Verified rose from 40% to 80%+ in a year.
Anthropic shows how to keep Claude coding for days: a list of 200+ tasks in JSON
Anthropic describes a harness for long-running Claude Agent SDK agents: an initializer agent, git, a progress file, and 200+ features the model may not edit.
Anthropic Reveals How It Isolates Claude: Users Simply Approved 93% of Confirmations
Anthropic detailed how it protects agents in claude.ai, Claude Code and Cowork: users approved 93% of requests, and a phishing prompt stole AWS keys in 24 of 25 attempts.