$ cat briefs/daily/2026-10-23.md
2026-10-23
October 23, 2026 (Fri)
3 stories · 2 paper picks · 2 new pages
Top Stories
1. Agent Skills moves to the Agentic AI Foundation
Anthropic is donating the standard's name, logo, GitHub organization and domain. It reports adoption by more than 70 agent platforms; users and contributors need not join the foundation.
Why it matters: portable agent procedures gain a shared governance home, while adoption counts alone do not establish reliability. → Agents (LLM Agents) (source)
2. Decisions API reaches general availability
The GPT-6 Luna interface returns bounded probabilities, choices or scores, with up to 400 questions per request and native GPT-Live integration. Dedicated endpoint rates are $0.10/M input · $0.01/M cached input, with no output-token charges.
Why it matters: applications can separate frequent routing decisions from generation; these endpoint rates do not replace Luna's ordinary API pricing. → Model Routing (source)
3. Agent monitors face attacks spanning contexts
An Anthropic Fellows study finds September configurations of Auto Mode and an approximation of Guardian miss over 55% of adversarial attacks in its aggregate. Attackers often know monitor internals; sandbox defenses were not tested.
Why it matters: per-action review can miss intent spread across sessions and compactions. These experimental rates do not measure ordinary-user incidents or current production safety. → Agent Runtime Containment (source)
Paper Picks
Cache engineering: direct KV cache control improves average benchmark score by 1.7 points while cutting average cost 27.7% against threshold-triggered summarization. Read it for the finding that identical token prefixes can hide different cache states. Hardware and full benchmark details are absent from the abstract. → Cache Engineering for Agents (source)
Agent4Sci: research skills raise Claude Opus's mean relative improvement from 10.4% to 20.1%, and Codex's from 4.4% to 8.1%, on scientific ML tasks. Read it for distinguishing engineering optimization from domain insight. These are author-reported relative improvements, not task-success rates; exact model versions are absent from the abstract. → Agent4Sci: Automating ML for Science (source)
Watch
New in Wiki
- Two research notes: Cache Engineering for Agents and Agent4Sci: Automating ML for Science, with methods and evidence limits. (cache) (science)
Updates
- MCP — Model Context Protocol records the proposed Agent Skills roadmap: versioned releases, a formal proposal process and security guidance. This is not an MCP protocol revision. (source)