Workflow Intel

Triage queue

Workflows

orchestration

2026 in LLMs (so far)

Simon Willison's Weblog (LLM / AI engineering) · Sep 27, 2026

INVESTModel: TRACK
ev2cred2dur2com1cmp2app3

Claim: 2026 has seen coding agents (Claude Code + Codex) mature from unreliable to production-ready; Claude Opus 4.5 and GPT-5.1 (Nov 2025) crossed an inflection point where agentic coding moved from "often make mistakes" to "reliable enough to use on a day-to-day basis".

Evidence: Author (credible AI observer) gives first-person account of model evals and practical deployment experience; anecdotal benchmark (pelican-on-bicycle SVG generation); notes 40+ conference sessions on agent security/sandboxing as signal of maturation; personal testimony of sustained project delivery with agents; mentions specific model releases (Claude Opus 4.5, GPT-5.1) timed to Nov 2025.

Rationale

First-person observational account from credible source (Simon Wardley, long-time analyst/speaker) about the exact dual agents (CC+Codex) Timothy runs. Direct applicability to monorepo workflow. Evidence is reproducible signal (personal deployment experience across multiple projects Jan–Jun 2026, conference attendance as independent corroboration of maturity trend). Durability moderate: depends on model capability but pattern of agentic coding + orchestration survives minor version bumps. No commoditization lever cited. Hype marker: "invisible line" is unquantified superlative framing, though claim itself is grounded in observed behavior. Verdict TRACK (not INVEST yet because claim is observational rather than prescriptive; worth monitoring but not actionable without roadmap/release specifics).

v2.1.170

Claude Code — Releases · Jun 9, 2026

INVESTModel: ADOPT-CHEAP
ev2cred3dur2com1cmp2app3

Claim: Claude Code v2.1.170 introduces Claude Fable 5 (Mythos-class model) with capabilities exceeding all prior general-availability models, and fixes a session-transcript bug in VS Code integration.

Evidence: Release notes from Anthropic official source (anthropic.com/news). Two concrete deliverables: (1) new model access gated behind version requirement; (2) named bug fix (sessions not saving transcripts from VS Code terminal with inherited CC environment variables).

Rationale

Timothy runs CC + Codex on TS/Next.js/Supabase; this is a direct version bump for his primary tool with a Mythos-class model (higher capability tier) and a bug fix (VS Code integration—directly in his stack). The bug fix is immediately applicable; new model access is a capability play. Composition_specificity=2 (not 3) because the release is primarily a maintenance bump + model tier upgrade, not a new orchestration pattern. Durability=2 (model capabilities are subject to next-release change, but the tooling fix persists). No commoditization signal cited. Hype marker flagged for the unquantified superlative on capability claims. Verdict: ADOPT-CHEAP because it's a straightforward version update with tangible bug fix + model upgrade, not a strategic investment decision.

v2.1.166

Claude Code — Releases · Jun 6, 2026

INVESTModel: ADOPT-CHEAP
ev2cred3dur2com0cmp2app2

Claim: v2.1.166 of Claude Code/Codex releases fallback model routing, thinking token controls, cross-session security hardening, and fixes for image processing, remote sessions, terminal rendering, and git worktree crashes.

Evidence: Official release notes listing ~21 discrete fixes and features, including fallbackModel configuration, MAX_THINKING_TOKENS controls, SendMessage auth hardening, and specific bug fixes (image processing token waste, remote session hangs, terminal flickering, git worktree crash-loops, PowerShell hanging, macOS CPU spinning).

Rationale

Standard v2.1.x maintenance release with incremental fixes (terminal flickering, git worktree crashes, image token waste, remote session stability) and minor feature additions (fallback model routing, thinking token controls). Fallback model routing and thinking controls are relevant to Timothy's CC/Codex loop and may reduce token waste when primary model is unavailable; remote session fixes improve reliability of the dev loop. However, these are routine maintenance patterns, not structural innovations. Applicability is moderate (fixes real pain points in the monorepo/terminal/session flow) but not high (no token-spend reduction mentioned, no capability shift). Composition_specificity is 2 because fixes are incremental and apply to the Claude Code tool itself, not to the monorepo/TS/Supabase/Vercel composition. Durability is 2: fallback routing and thinking controls may be overwritten by future model behavior shifts, but terminal/session stability fixes are durable. No commoditization signal (these are maintenance, not forward-looking integrations). Source remains at baseline credibility 3 — this is official release notes with concrete, verifiable changes.

INVESTModel: ARCHIVE
ev2cred3dur3com0cmp2app2

Claim: OpenAI frontier models and Codex are now generally available on AWS, enabling enterprises to use them through AWS environments and procurement workflows.

Evidence: Official product availability announcement stating GA (general availability) status on AWS, mentioning integration with AWS environments, controls, and procurement workflows.

Rationale

Credible announcement of a real distribution channel; Timothy's Codex usage (monorepo context) gains AWS pathway optionality but does not fundamentally change his CC+Codex workflow architecture or token efficiency. Durability is high (this is a stable commercial partnership, not a model-specific workaround). Composition_specificity is moderate—it's procurement flexibility, not a stack-defining pattern. Applicability is moderate because it offers an alternative environment (relevant if Timothy operates in AWS), but does not directly cut token spend or reshape the CC/Codex loop. No hype markers detected; routine product language.