The cheapest win in agents this week was a router, not a bigger model

The 60-second catch-up

OpenAI published the numbers on AI doing its own research, and its chief scientist published the warning on the same day.

By mid-August the research organisation was running 3.1 agent-workdays for every human workday, and the median researcher was burning more than $600 a day of inference at API prices (OpenAI). Hours later, chief scientist Jakub Pachocki argued that this feeds recursive self-improvement while alignment and monitoring lag, and drew a distinction worth stealing: goal alignment asks whether the agent pursues the objective, value alignment asks whether your constraints survive when the objective gets hard (OpenAI).

The headline number is 3.1. The number that actually matters is buried: more than half of successful four-to-eight-hour agent tasks still needed at least one human intervention. The automation is real and it is supervised, and the supervision is now the job.

→ do this: if you report agent adoption as "tasks started", change it to "tasks that shipped with no intervention". That second number is the one that moves, and it is the one that tells you whether you have leverage or a very fast intern.

Three unrelated forces landed on the pricing page in the same week.

Jason Fried came out publicly against per-seat pricing, on the grounds that per-user billing never mapped to value (Jason Fried). Kyle Poyar published data on whether AI assistants can even read a pricing page well enough to quote it (Kyle Poyar). And a study found Google's AI Mode surfaces the same products 21.6% more expensive than traditional search does (Productrise). Meanwhile California's SB 122 puts 8-10% on top of the software bill (SaaStr).

Seats stop making sense when agents do the work. An intermediary you cannot see is choosing what to show. A tax arrives regardless.

→ do this: paste your pricing page into an AI assistant and ask it what one seat costs and what is included. If it cannot answer cleanly, you have found your Q4 project, and it is not a redesign.

The cheapest win in agents this week was routing, not a bigger model.

Spotify reported roughly 90% mean token savings by pushing large file reads and predictable generation from Claude Code to cheaper workers, keeping the frontier model for debugging and architecture (Spotify Engineering). GitHub is making the same bet at the product layer with multi-model orchestration (GitHub). And a new observability tool exists specifically to hard-block the agents burning your budget.

→ do this: measure spend per agent run before you touch a prompt. Most teams tuning prompts for cost have never looked at which calls actually cost them anything.

Tools

Build & ship

• ⭐ cn — Drop-in replacement for clsx and tailwind-merge that keeps the same APIs. Zero dependencies, and it benchmarks its most common component call at 30× faster against a corpus of 144,265 real calls harvested from 58 open-source codebases. · Open source (MIT)
• Vidact — Compiles a documented subset of React 19 into direct DOM operations, so component bodies run once and a state change touches only the nodes that depend on it. Rejects unsupported APIs at build time rather than silently falling back to the React runtime. · Open source (MIT), beta

AI & agents

• ⭐ Traccia — Traces every agent's LLM calls and hard-blocks the ones breaking budget or policy, then exports the evidence for an audit. · Free tier · $99/mo
• TaskShell — Connects to Claude, ChatGPT or Cursor over MCP, so you tell the agent what you finished and it ticks off the task and subtasks. · Free trial, then paid

Design & create

• Causal — An AI canvas workspace for working ideas up into something shaped. · Free plan (no expiry, no card) + paid

Growth & ops

• Lightfield — An AI-native CRM that reads your email, calendar and calls to keep records current, then preps the next meeting and drafts the follow-up. · 5,000 free credits · Pro $1,000/mo
• Pluto — A voice agent takes your career story in about ten minutes and turns it into a profile recruiters and AI agents can search, then makes the introduction when something fits. · Free for professionals
• Sandcastles — Short-form video and script generation, with an MCP endpoint so an agent can drive it. · From $39/mo

Resources

Steal the template

• How to Get Your First Customer (Then Your First 3) — Noah Kagan, on the part everyone skips because it does not scale.
• Interface Cheat Sheet — a monthly interface pattern library with interactive demos and source. $7.99/mo.

Learn the craft

• Using AI for UX Work: Study Guide — NN/g, structured rather than listicle.
• Behind the Build: Generative Plugins and Shaders at Figma — how the feature was actually built.
• How We Raised Mobile End-to-End Test Stability to 98% — Shopify rebuilt around a strict builder API that demands an assertion after every action and labels the escape hatches UNSAFE_. From 50% to 98%.
• Reducing Zod's Memory Footprint by an Order of Magnitude — method memoization, written up properly.

Benchmarks you can re-run

• LLM-as-a-Verifier — a general-purpose framework for verification signals.
• Random Attention — Salesforce AI Research's code for its KV cache eviction paper.
• GPT-6 Astra in Code Review: Gains, Privacy, and Cost — CodeRabbit's own evaluation, with the cost column included.
• Google AI Mode shows the same products 21.6% more expensive — the method is in the post, so you can run it on your own category.

Reads

☕ Under 5 minutes

• My View on Design Engineering — Jim Nielsen.
• Goodbye per-seat pricing — Jason Fried on why per-user billing always rubbed him the wrong way.
• No one wants to hear this, but there's no magic distribution channel — Dylan Txa, who took GojiberryAI to over $4M ARR in a year, working one channel at a time by MRR band.
• Concrete, Silicon, & Leverage — Tomasz Tunguz on the data-centre debt wave.
• You Get What You Reward.
• AI Handles Incidents, Engineers Lose Touch with Their Systems — Sylvain Kalache on comprehension debt.

🍵 5–10 minutes

• On the YC S26 batch — Return of the Atoms — Jared Heyman, who has been to nearly every Demo Day since his own in 2013.
• The Point of a System Is Fewer Decisions — Anton Sten. If execution is cheap, a design system's value is the decisions it stops you re-litigating.
• AI Safety Is Not the Same as Security — Martin Alderson argues the frontier labs have conflated them.
• How China is turning AI tokens into everyday consumer rewards — Rest of World. Tokens bundled with coffee, credit cards and dumplings.
• Portal by Spotify Cut My Claude Code Token Usage by 90% — the routing write-up behind the catch-up above.
• I've factored the RSA keys of a Certificate Authority from the 90s.
• Towards Self-Driving Codebases — more agents produced more code without producing better software. What is actually missing.
• When the canvas starts acting, who's really in control? — what designers still control once the interface has agency.

📚 Longer, worth it

• Is your pricing readable by AI? — Kyle Poyar, with the data.
• California Is Taxing SaaS and AI Tools — what SB 122 does to buyers and to vendors, separately.
• An Alien Mind — Jakub Pachocki's own argument, at length.
• OpenAI and the Wiki Incident — Zvi Mowshowitz.
• An Interview with OpenAI President Greg Brockman About Astra and Alignment — Ben Thompson. Long, and the alignment section earns it.

Numbers

• 3.1 — agent-workdays per human workday inside OpenAI's research organisation by mid-August (OpenAI).
• More than half — successful four-to-eight-hour agent tasks that still required at least one human intervention (OpenAI).
• $600+ a day — median OpenAI researcher's inference consumption at API prices (OpenAI).
• ~90% — mean token saving Spotify measured across four Java monorepo scenarios by routing work away from the frontier model (Spotify Engineering).
• 50% → 98% — Shopify's mobile end-to-end test stability, before and after the rebuild (Shopify Engineering).
• 21.6% — how much more expensive the same products are when surfaced by Google AI Mode versus traditional search (Productrise).
• 8–10% — the software cost increase California's SB 122 implies for buyers (SaaStr).
• $4M — GojiberryAI's ARR one year in, reached channel by channel (Dylan Txa).
• 144,265 — real cn() calls harvested from 58 open-source codebases to benchmark the library against its predecessors (cn, MIT).
• $103B — Nscale's contracted revenue as pitched to investors, up from $51B before its $45B compute deal with Anthropic, on contracts averaging 5.7 years (The Information).

What happened

Models & math

• Anthropic published a formalization of Fermat's Last Theorem (Anthropic).
• Robocurve published an evaluation of GPT-6 Astra on robotic manipulation, reporting where it wins and where it does not (Robocurve).
• OpenAI's headline AGI benchmark number came from the harness rather than the model (The Next Web).
• Extropic published Z1T, transformer-like models built for its Z1 chip (Extropic).

Safety & policy

• OpenAI's chief scientist Jakub Pachocki warned that recursive self-improvement is plausible while alignment and monitoring lag, and that chain-of-thought monitoring may narrow as models improve (OpenAI).
• Every binding AI review Washington has proposed has come back voluntary (The Next Web).
• California's SB 122 extends sales tax to SaaS and AI tools, landing 8-10% on software bills (SaaStr).

Business moved

• Anthropic's IPO launch shifted toward mid-October (CNBC).
• Adobe named Anil Chakravarthy chief executive, replacing Shantanu Narayen (CNBC).
• Nscale told prospective investors it holds roughly $103B in contracted revenue ahead of a possible September IPO, roughly double its position before a $45B compute agreement with Anthropic (The Information).
• Jason Fried publicly called time on per-seat pricing, arguing per-user billing never mapped to value (Jason Fried).

Platforms shifted

• Runway introduced Solaris (Runway).
• Microsoft announced Project Zenith, a ready-to-code Windows experience for developer-class devices (Windows Developer Blog).
• GitHub shipped Project HydraFusion, aiming at frontier quality through multi-model orchestration (GitHub).
• Google continued turning the Gemini desktop app into a superapp (TestingCatalog).

We skipped a handful of stories this week that were interesting but had nothing in them for someone building — a Disney auction, a Harry Potter logo argument, and a very good 26-minute piece on WebAssembly proof-of-work that we could not get past its own bot wall to check.