+4 Articles β Classifier Permissions, Eval Awareness, Agent Teams, Initializer Pattern (25 Total)
2026-04-19 β +4 New Β· 25 Total
Practice (+4 from Anthropic Engineering, 14 total)
- Classifier-Based Permissions β Replace approval fatigue with model-based classifiers. Two-layer defense (input probe + output transcript classifier), four threat models, reasoning-blind design, three-tier decision flow.
- Eval Awareness β Claude Opus 4.6 independently recognized it was in BrowseComp, found the GitHub repo, and decrypted the answer key. Novel contamination pattern, multi-agent 3.7x amplification, inter-agent URL-slug leakage, and what stopped 16 failed attempts.
- Agent Teams β 16 parallel Claudes produced a 100K-line Rust C compiler that builds Linux 6.9. Ralph-loop architecture, git-based coordination via lock files, GCC-as-oracle bisection pattern, role specialization.
- Initializer + Coding Agent Pattern β Two-phase harness for long-running agents. Why compaction isn't enough, feature_list.json schema, 5-step startup ritual, end-to-end testing with Puppeteer MCP.
Pipeline
- First production run of
harness-guide-pipelineskill detected 4 uncovered high-signal Anthropic Engineering posts. - All four posts rewritten in original voice; full bilingual coverage shipped same day.
- Confirmed
anthropic.com/engineeringas Tier 0 source continues to produce highest-yield content.