AI builder notes - September 2026

Notes from my September liked tweets and GitHub stars. Three things stood out this month: using deterministic CPU/React counters so agents can optimize performance without noisy benchmarks, enforcing token limits in hooks instead of skills, and the wave of tiny sub-100ms “System One” decision models.

Main takeaways from September

Give perf agents deterministic counters, not wall-clock timers. Anthropic made claude.ai 3x faster in 2 weeks, merging 3,000+ changes across 150+ concurrent agent threads [1]. Wall-clock millisecond benchmarks on shared CI runners are way too noisy for an agent loop to hill-climb. Instead, they measured deterministic Valgrind instruction counts on the backend and React commit counts on the frontend, with a CI ratchet so no PR could increase the count.
“The hook enforces, the skill suggests.” Spotify shared a great writeup on their internal coding agent setup [2]. Even when SKILL.md explicitly told Claude Code not to read huge files raw, agents kept doing it anyway. So they built a shunt pre-tool hook that intercepts large file reads and routes them through a cheap summarizer model, cutting Claude Code token usage by 90%. Also cool: Spotify deleted their early PR judge subagent once base models improved from 25% -> 80% task success on their own.
“System One” decision models (Bool, Choice, Score) are insanely fast for routing and classification. TypeSafe launched Jev [3] [4] - a non-generative decision API trained via RLCD that returns typed booleans, choices, or calibrated probabilities 20-200x faster than chat models, with output tokens free and up to ~255 options scored in parallel [5].
Fast element picking != long-horizon planning. Good reality check from @gregpr07: a decision-only browser agent booked a flight in 7 seconds for $0.0039 [6], which looks amazing in a demo. But across 20 harder multi-step web tasks, it scored 1/20 vs 17/20 for a full BrowserCode + Luna agent coz recovery and state tracking still need a reasoning loop [7]. Use System One for fast inner-loop guards and routing, not as the whole brain.
Formal methods + coding agents actually work now. Boris Cherny shared how they used the Agent SDK with Lean and TLA+ to formally model concurrent state transitions, finding real race conditions and landing 16 fixes. [8]

Cool uses of fast decision models

Semantic SQL inside DuckDB: running a fast boolean decision model directly inside a DuckDB WHERE clause over 50,000 support tickets in Parquet to filter for genuinely angry customers that keyword LIKE queries miss. [9]
Live voice call outcome forecasting: @muratcan ran 38,012 turn-level call outcome predictions across 2,029 receptionist traces at 118ms median latency, plus 10,145 post-call QA checks in 26 seconds for $3. [10]
Bulk SEO internal-linking audit: @borjafat scored 8,790 candidate link pairs across 586 pages in 45 seconds for $0.21. [11]
Voice router for terminal panes: @jobelshaji95 paired MediaPipe + local STT + a typed Choice call to classify in milliseconds whether spoken audio is a window command for herdr or dictation meant for the active coding agent. [12]

Production guardrails & kernel experiments

Cursor Rollouts & Security Reviewer: Cursor added pre-merge monitoring plans, post-deploy metric checks that open auto-revert PRs on regressions, and whole-codebase taint tracking that pushed security review comment acceptance to 60-70%. They also trimmed their base system prompt by 66% and moved MCP tools to on-demand loading (46.9% fewer tokens). [13]
Perplexity’s Lily on Apple Silicon: Denis Yarats shared ablations from their custom Metal kernels + Rust runtime showing that once direct single-stream decoding is properly memory-bound tuned on Apple Silicon, speculative decoding actually lost on throughput. Always benchmark your hardware before adding complexity. [14]

Starred GitHub repos from September

headroomlabs-ai/headroom - local reversible compression proxy (headroom wrap claude|codex) that sits between your CLI and the API, shrinking huge JSON tool outputs, logs, and file reads while keeping your earlier KV-cache prefix untouched
redhat-et/ripwire - zero-dependency C++23 CLI (ripwire . --for="...") that indexes a repo in ~0.3s and hands the agent a ranked call graph, blast radius, and exact tests to run in ~5% of the tokens of a naive grep loop
AltanS/collie - mobile web UI served over tailscale serve so you can check on your herdr, tmux, or zellij coding agents from your phone, get push notifications when a pane blocks on approval, and tap a quick reply
ayghri/i-have-adhd - surprisingly effective SKILL.md that forces agents to put the exact command/answer on line 1, cap bullet lists at 5 items, and skip all the polite filler
google-deepmind/science-skills - DeepMind’s SKILL.md bundle for genomics, AlphaGenome, UniProt, and scientific literature search
Rigellute/spotify-tui - lightweight Rust Spotify client right in the terminal (spt)