Factory Review · 2026-W45

What the factory watched this week.

Every week our pipeline scrapes the model catalogs and vendor blogs, then judges each item against one question: would this actually improve something we run in production? Verdicts below. Watch = interesting but unproven claims · Adopt = earned a place · Ignore = noise · Product Input = a feature idea, not a factory change.

30 items reviewed: 22 dropped (mostly Tier 3 generic agent-building/vibe-coding checklists and model-watch catalog-churn variants already covered by existing doctrine, architecture, or the locked Model Allocation policy), 7 held at Watch, 1 at Factory Candidate. The one Factory Candidate (, Simon Willison relaying the Claude Code team on context-first vs. restrictive prompting) directly challenges the prohibition-heavy style used across CLAUDE.md/AGENTS.md/skills, with a cheap bounded single-skill test proposed. Everything else held at Watch pending either direct-doc verification (Opus 5 safety-fallback claims, a reported 'Managed Agents' feature) or a concrete triggering need (self-hosted forensic model, cache-aware routing costing, page-context token registry, sycophancy probe) that hasn't yet materialized in this portfolio. No Immediate Factory Upgrade or Product Input items this week — nothing met the materiality bar of naming a specific pending product decision.

The headline calls

Watchanthropic-news

claude opus 5

Opus 5 reportedly allows mid-conversation tool-list changes without invalidating the prompt cache, plus automatic safety fallbacks when a request trips a refusal — both would change how agentic harnesses structure tool definitions and handle refusal states.

Why this verdict: Tier 1 source (anthropic.com/news), relevant given the recent judge commit that mistreated a safety refusal as a format problem, but the mechanism is a vendor summary we haven't verified against docs — hold at Watch.

claims to check before believing:
  • Opus 5 'automatic safety fallback' behavior and mid-conversation tool-change cache semantics as described
Watchlatent-space · AINews

"Laguna S 2.1 Released: Cheaper than Deepseek v4 Flash, Better than V4 Pro"

A reported Anthropic 'Managed Agents' feature adds per-agent effort controls, a 500-skills-per-session ceiling, and sub-agent event streaming, which would interact with the Model Allocation effort policy and any skill-heavy session.

Why this verdict: Tier 3 aggregator relaying what reads like a Tier 1 fact; the feature name and limits are unverified — holds at Watch pending direct confirmation.

claims to check before believing:
  • Existence and exact specs of an Anthropic 'Managed Agents' product with per-agent effort controls, 500 skills/session cap, and sub-agent event streaming

The rest of the week (6)

Factory CandidateA Fireside Chat with Cat and Thariq from the Claude Code team
WatchSecurity incident disclosure — July 2026
Watchseer agent page context
WatchModel Routing Is Simple. Until It Isn’t.
WatchI'm still in control, but I'm not coding
WatchControlling Reasoning Effort in LLMs