Anthropic's announcement of Fable 5.1 and Mythos 5.1 — one base model, two guardrail postures.
- official
- models
- announcement
- fable
Which Tools Do Coding Agents Install? (Armature)
16,893 sessions of experiment data on how Claude Code, Codex, and Cursor actually choose third-party tools.
Armature's experiment on tool selection by coding agents: 16,893 sessions, 1,163 prompt variants, 75 repos, 10 languages (first public wave: 5,292 valid sessions). Headline findings: the three agents agreed on the same tool in only 42% of scenarios; Claude Code leans on prior knowledge (~30% web search) while Codex searched 94% of the time; "being mentioned is not being chosen" — PayPal was cited 139 times and won zero selections (Stripe won 124); winners differ by language for the same need (Resend-TS, SendGrid-Python, Postmark-Go). The author discloses Armature sells dev-tools growth services — read with that in mind.
Charted
Related on the map
Anthropic's canonical essay on agent design — workflows vs agents, and five composable patterns.
- official
- agents
- engineering
- patterns
12-Factor Agents
GuideHumanLayer's 12 design principles for production-grade LLM agents — the closest thing to a 'best practices' manifesto.
- agents
- best-practices
- production
- principles