⚠️ Sponsorship
Two paid placements: Augment Code (Agentic SDLC report on coordinated agent teams across the six stages of software delivery) and Airbyte (Context Store pre-indexing for Claude Code — 50+ connectors into one queryable store). Granola also appeared as a "Presented by" in the Signals section. Editorial items (Anthropic security plugin, Cursor Router, Managed Agents) are organic coverage.
Issue contents
Top News — Anthropic Claude Code Security Plugin (beta)
Multi-agent vulnerability scanner running inside Claude Code terminal. Three-layer approach: (1) per-file-edit pattern scan flagging eval(), os.system(), innerHTML, hardcoded secrets with zero added cost; (2) post-turn git diff review for logic-level flaws — auth bypasses, insecure direct object references; (3) pre-commit deep cross-file review catching issues that only show up in context. Free for all Claude Code users. Install: /plugins in terminal → /reload-plugins. Reported to have caught 500+ previously unknown high-severity issues in open-source codebases during internal testing and competitions. Powered by Opus 4.6 for deep scan mode.
Top News — Cursor Model Router
Auto-routes each Cursor request to the best-fit model by task complexity instead of user manually choosing. Three configurable modes: Intelligence (best quality, still cheaper than always using frontier), Balance (solid quality at ~36% lower cost), Cost (maximum savings for light tasks). Early testing with thousands-of-user teams showed 30–50% cost reduction vs. routing everything to Opus 4.8 with no quality drop. Cost-per-commit metric was also lower. Live on Teams and Enterprise plans; admins can lock model sets and defaults per team. Cursor's Auto mode is now powered by Router by default.
Top Repo — Claude Managed Agents: 500 Skills, Effort Controls, Webhooks
Batch of capability upgrades to Claude Managed Agents API:
- Session seeding: up to 50 events in a single session create call (was a two-call create-then-seed flow)
- Effort controls: per-agent effort level (Low/Medium/High/Max) set in the model config — lower effort = faster + cheaper, higher = more thorough
- 500 skills per session: up from a much smaller limit; skills are task-specific instruction sets stacked across all agents in a session
- Webhooks: lifecycle notifications for environments and memory stores — no more polling
- Sub-agent event streaming: watch child agent actions in real time
Available via anthropics/claude-cookbooks on GitHub.
Signals (selected)
- Fast tokenizer: open-source tokenizer 1000x faster than HuggingFace, hitting 24 GB/s
- Outlines: open-source library that forces LLMs to output valid structured data every time (15k+ stars)
- Schmidhuber survey: comprehensive review of how modern AI agents learn to improve themselves
- Microsoft prompt rewriter: open-sources a tool that rewrites and improves its own AI prompts
Why this is in the vault
Three Anthropic/Claude platform items in one issue, all directly relevant to RDCO's operating infrastructure: security hardening for Claude Code (which runs continuously in the RDCO harness), model routing cost patterns (relevant to multi-agent dispatch decisions), and expanded Claude Managed Agents capacity (500 skills, effort controls, webhooks) that maps almost exactly onto the RDCO skill-layered agent architecture. The Managed Agents item is the most load-bearing; it closes open architectural questions around polling and session complexity.
Mapping against Ray Data Co
The 500-skills-per-session cap in Claude Managed Agents is the sharpest signal: RDCO's skill library sits in ~/.claude/skills/ with ~50+ named skills, well within range for a single session. That means the effort-control and webhook additions are immediately actionable without architecture rework — worth queuing as a harness upgrade candidate. Specifically, the webhook-for-memory-store lifecycle change could replace the manual cron polling that currently drives parts of the autonomous agent loop (process-inbox, sync-contacts, open-threads-check all poll; webhooks would let those respond to state changes instead).
The Claude Code security plugin is the lowest-friction win. RDCO runs Claude Code for code generation, vault writes, and deploy scripts in Claude Code sessions. Three-layer scanning (pattern → diff → cross-file) fires automatically at edit and commit time — no workflow change needed. Also relevant to the phData DSA role, where code quality in client delivery matters and a security catch at commit time reduces rework risk.
Cursor Router is lower priority for RDCO since the primary coding tool is Claude Code not Cursor. But the underlying pattern — routing by task complexity to tier model costs — is architecturally relevant: Managed Agents' new effort controls are essentially the same idea applied to multi-agent orchestration. That alignment is worth noting for future harness decisions around which agents get High vs. Low effort.
Related
[[2026-07-22-alphasignal-claude-skill-recording-imo-exploitgym]]— Claude skill recording and expanding per-agent capability surface; the 500-skill cap is the capacity unlock that makes skill-recording practical at scale[[2026-07-21-alphasignal-llm-fingerprinting-cursor-agent-costs]]— previous Cursor cost signal; Cursor Router is the shipped answer to the cost problem surfaced there[[2026-07-21-technically-harness-engineering]]— harness engineering context; effort controls and session-seeding are new harness primitives worth folding in[[2026-07-21-innermost-loop-harness-as-generalizer-sandbox-escape]]— sub-agent event streaming and webhook hooks bear directly on the innermost loop and sandbox-escape architecture discussed there