Today's haul: 6 tools · 11 resources · 23 reads · 10 numbers · 15 things that happened. Every tool, resource and read below has a working link. No link, no listing.
⚡ 60-Second Catch-Up
Open weights crossed the halfway line somewhere real, and the frontier labs answered on price. Open-source models went from 28% to 62% of tokens served at Vercel over two months, a chart from Vercel's Guillermo Rauch that Gavin Baker put numbers around — and he thinks the end state is closed frontier tokens holding 60–90% of economic value on only 15–25% of tokens. Four days earlier OpenAI cut GPT-5.6 Sol's API price by more than 20% for three months, input from $5 to $4 per million and output from $30 to $20. Meanwhile Anthropic's cheaper Opus 5 overtook Fable 5 in corporate spending within a month of launch, and one developer handed Qwen 3.8 27B a reverse-engineering job he assumed needed a frontier model and got it back in 30 minutes. → The interesting number is 62%, not the price cut. Token share moves when the cheap option clears the bar on ordinary work, and ordinary work is most of what you run. Pick your three highest-volume tasks this week and run each one against an open model and your current default, then grade them yourself — if you can't tell the difference, you have been paying frontier prices for commodity output.
The agent harness stopped being plumbing and became the product. Three open-source harnesses landed in one day: Headlong, a Laude/MIT microharness whose agent never sleeps and generates its own thoughts between messages in under 10K lines of Bash; Apache Maka, a local-first agent workspace that records every model message, tool call, permission decision and termination event as an append-only log; and Proliferate, an open-source IDE that runs Claude Code, Codex and OpenCode in parallel isolated workspaces. Dan McAteer's evolution-of-the-harness essay argues the whole game has shifted to optimising human attention rather than model capability. And the practitioner end looks like Ryan Carson, who spent $20,000 on Devin in a month, runs 10–15 threads at once and ships up to 40 pull requests a day — while keeping a handwritten priority list on paper beside him. → Note what all four have in common: the constraint is you, not the model. Maka's append-only log and Carson's paper list are the same idea — an agent you cannot audit or prioritise is an agent you cannot delegate to. Before adding a fifth tool, write down how you would answer "what did it do and why" for the four you already run.
Your own website quietly became the thing that gets quoted. Reddit's share of ChatGPT citations fell from 3.8% to 0.5% in a single week — an 86% collapse — after ChatGPT changed how it searches to look at official sites first, a shift documented by marketer Tom Orbach. Ask it what Notion costs now and it quotes Notion's pricing page, not a thread about Notion. → This is a content brief, not a trend. Three pages decide whether you get quoted: a pricing page with real numbers, because nothing can quote "contact us"; a "you vs competitor" page, because buyers ask AI to compare and it quotes whoever bothered to write one; and an FAQ that answers each question in its first sentence. If you have been treating your marketing site as a brochure, it is now an API.
🛠 TOOLS — 6
AI & agents
Headlong ⭐ — An agent microharness built on persistent agency: the agent keeps thinking between external interactions in a self-guided loop, rather than freezing until the next request. A message from a human doesn't start a session — it's one more observation landing in an ongoing thought stream. Core is under 10K lines of Bash. A Laude Institute / MIT collaboration, announced this week · Open source (Apache-2.0)
Apache Maka ⭐ — A local-first AI agent workspace where model messages, tool calls, tool results, permission decisions and termination events are all recorded as an append-only log — so session data stays on your machine and every action stays auditable after the fact. Currently an Apache incubating project · Open source (Apache-2.0)
Proliferate — An open-source AI IDE that runs Claude Code, Codex, OpenCode and others in parallel inside isolated workspaces, locally or in the cloud. Supports self-hosting and reusable workflows, so a task you've run once becomes a thing you can re-run rather than re-prompt · Open source (AGPL-3.0)
Grok Bot — Text-thread agents you message like a teammate, run many at once, and drop into a group chat so they hand work between themselves without you in the middle. Left beta on August 21 and is now bundled rather than sold separately · Included with SuperGrok Plus, SuperGrok Heavy, Cursor Pro+, Cursor Ultra and Cursor Teams (Standard and Premium)
Design & create
Balsa UI — An agent-native, open-code component system for Vue and React. Rather than shipping a black-box package, its registry hands your coding agent the actual source so it can extend and evolve the components — a contract between a design system and the agents editing it. Starts with npx balsa-ui@latest init · Free (installs via npx; no pricing page)
Taku — A marketplace that turns published AI apps, agents and workflows into one-click tools you can run without installing anything, then assembles several into a "stack" from a plain-language description of your goal. Creators package their own setups as shareable, remixable bundles · Free to start
📚 RESOURCES — 11
Steal the template
How Ryan Carson Manages 40 PRs a Day with Devin and Codex — the full workflow behind the $20,000 month: 10–15 Devin threads sorted into P0/P1/P2/Bugs folders, treated the way a manager treats direct reports. Devin for cloud work, Codex when the frontend needs close attention, and the reasoning for the split is latency and browser access, not loyalty.
Automate Business Health Checks with a Watchdog AI Agent Playbook — Carson's Watchdog reviews every customer account, pulls recent activity and errors from Sentry and internal logs, picks the three most important problems, and checks whether a recent PR already fixed each one. A customer-success function as one reusable agent thread.
Create a Production-Ready Design System Using Claude and Codex — Claire Vo breaks a Figma file into design tokens and a design.md spec with Claude, then hands that spec to Codex to build shared components across a monorepo. The split is deliberate: one model translates a visual system into a portable spec, the other turns the spec into interconnected code.
On teaching AI how you work — Karl Koch's eleven skills consolidated into one repository, so design intent survives from one session to the next instead of being re-explained every time.
Building a 24/7 Multi-Agent System: The SpaceXAI Playbook — a 100-minute architecture document mapping the path from one bot to an always-on team: explicit ownership, reusable skills, event-driven routines, typed handoffs, verification rules and approval boundaries.
Learn the craft
The AI-Native SDLC playbook — 47 minutes on the gap this year actually opened: AI made writing code fast while the process around it stayed the same shape, so the bottleneck moved somewhere nobody was measuring.
How to close $100K–$1M deals, step by step — Jen Abel on founder-led enterprise sales, at the deal sizes where the founder cannot yet hand it off and the playbook has to be personal.
The New MCP Roadmap — the five priority areas for upcoming spec releases, worth reading before you build anything load-bearing on the current version.
Benchmarks you can re-run
My Hands-On Review of GrokBot, Cursor Origin, and the Grok 4.6 Model — Claire Vo runs her evaluations blind, grades outputs herself and weights her own judgment at 70% of the final score. Grok 4.6 finished level with GPT-5.6 Sol at the top, ahead of Sonnet 5 and Opus 5 — though she still reaches for Sonnet 5 when she wants an agent that is good to talk to. The method is the transferable part.
📖 READS — 23
☕ Under 5 minutes
How Universities Should Prepare Founders — Paul Graham's answer is to teach no "entrepreneurship" at all: teach the building subjects well, show students real founders so it feels possible, then give them free time and leave them alone. Microsoft and Meta both started during Harvard's reading period — the one stretch when students are on campus with nothing due tomorrow.
Vibe Coding and Quality — Ilya Birman shipped a Mac app with Codex that beat the human-made alternatives inside a day, and comes out somewhere uncomfortable: that exacting quality may now be more reachable this way, not less.
A Sloppy Interface is a Security Liability — if AI makes faking a polished interface trivial, then craft that is expensive to imitate — micro-interactions, loading behaviour, edge-case handling — starts working as a defence, not just a brand asset.
The summer of open weights — Martin Alderson's read on the tipping point: aggressive pricing moves, a real field of competent open models, and the advantage now sitting with whoever serves them most efficiently.
How I Find Problems to Solve as a Staff Engineer — the senior-IC skill nobody writes job specs for: listen for the complaints, let them accumulate, notice when three unrelated ones are the same problem.
Figma is the New Dreamweaver — static artboards and Auto Layout hide responsive behaviour, loading performance, accessibility and unpredictable content, and the gap widens as products get more adaptive.
Fences, Not Sandboxes — Steve Yegge on the wrong mental model for constraining agents, and what to use instead.
The Evolution of the Agent Harness — how models absorbed the harness's capabilities, and why the remaining job of a harness is optimising human attention rather than model reach.
How Europe is killing makers and micro-entrepreneurs — compliance regimes written for factories, applied to people selling a hundred units, and the arithmetic of what that does to a one-person hardware business.
One AI Output is an Example, Not an Evaluation — Nielsen Norman Group on why a single good output proves nothing: you need repeated runs across representative inputs, averages with confidence intervals, and attention to how a system fails rather than only how often.
How Complex Systems Fail — Richard Cook's eighteen theses, still the clearest short statement of why single-cause post-mortems are usually wrong.
AI Chip Architectures — 110 minutes across NVIDIA GPUs, Google TPUs, AMD Instinct and Cerebras' wafer-scale engine, and what each one's design is actually optimising for.
Why your local LLM feels dumber than it is — quantisation, configuration and hardware variation, tested rather than asserted. If you concluded a local model was weak, this is the piece that says measure again.
Everything I own, owned — how easily commodity peripherals gave up microphones and webcams once someone bothered to reverse-engineer them.
Measuring benchmark optimization in speech recognition — three tests for detecting "benchmaxxing", plus real instances of models reproducing dataset errors in VoxPopuli and LibriSpeech. The method generalises past speech.
📊 NUMBERS — 10
28% → 62%. Open-source models' share of tokens served at Vercel, over two months — a chart from Vercel's Guillermo Rauch, with the figures set out by investor Gavin Baker.
3.8% → 0.5%. Reddit's share of everything ChatGPT cites, in one week — an 86% collapse, after ChatGPT began looking at official sites first (Tom Orbach).
$5 → $4 and $30 → $20. GPT-5.6 Sol's API price per million input and output tokens, cut for three months from August 21 (OpenAI).
$13 billion. The valuation at which Hugging Face reportedly gauged buyer interest — nearly triple its 2023 mark, with no agreement reached (RuntimeWire).
15%+. The increase in AI server prices expected next year as Nvidia passes rising memory costs to customers (StreetSignal).
200+. Employees laid off across Apple's Vision Pro, Siri and software engineering teams, alongside a scaling back of Vision Pro gaming and immersive video (9to5Mac).
27 billion parameters. The size of Faraday, the agent from DeepMind-alumni startup Inherent that outperformed larger Anthropic and OpenAI models at replicating research papers (TechCrunch).
40 pull requests a day. What solo founder Ryan Carson ships running 10–15 agent threads at once, at a cost of $20,000 in Devin in a single month.
1 million customers in two and a half years, zero advertising spend. Monzo's growth before its founder left, in Tom Blomfield's own account of it.
~11%. The share of spending on Anthropic tools that Fable 5 plateaued at, before the cheaper Opus 5 overtook it within a month of launch (Implicator.ai, on Ramp data).
📰 WHAT HAPPENED — 15
Models & math
DeepSeek released V4-Flash-Vision-Exp, an experimental multimodal model that adds image understanding — screenshots, diagrams, text extraction — and nearly matches Opus 4.8 on agent benchmarks. It works against OpenAI's Chat Completions and Responses APIs and Anthropic's Messages endpoint (The Decoder).
OpenAI cut GPT-5.6 Sol's API pricing by more than 20% for three months, covering pay-as-you-go API, Codex credits and eligible ChatGPT Work plans. Pro, Plus and Business subscription limits are unchanged (OpenAI).
Anthropic's cheaper Opus 5 overtook Fable 5 in corporate model spending within a month of launch, as low switching costs let companies route routine work down and reserve premium models for sustained autonomy (Implicator.ai).
Thomson Reuters launched its own frontier model, built on its internal data assets rather than licensed from a lab (Thomson Reuters).
Safety & policy
Anthropic made Claude Mythos 5 available for code scanning in Claude Security — widening access to what the model finds without letting most people prompt it. Users of partner tools get suggested patches or an alert, with no route to make it write an exploit (TNW).
The Model Context Protocol published a new roadmap setting five priority areas for coming spec releases (MCP).
Business moved
Hugging Face worked with a bank to gauge buyer interest at $13 billion or more, nearly triple its 2023 valuation, on the strength of its model hub and developer ecosystem. No agreement had been reached (RuntimeWire).
Apple laid off more than 200 people across Vision Pro, Siri and software engineering, scaling back Vision Pro gaming and immersive video while reorganising Siri around its new AI platform (9to5Mac).
Meta hired Luke Metz from OpenAI — the second time he has left, having gone to Thinking Machines in 2024 and rejoined OpenAI earlier this year (Axios).
Nvidia plans to pass rising memory costs to customers, with AI server prices set to rise more than 15% for systems shipping next year (StreetSignal).
Outer Biosciences is training models on living human skin, keeping donated tissue alive, measuring how it responds to compounds and training on the resulting biological data (TechCrunch).
Inherent, founded by DeepMind alumni, said its agent Faraday beat larger Anthropic and OpenAI models at replicating research papers, using 27 billion parameters (TechCrunch).
Platforms shifted
Slack launched Slack Code, letting teams tag Claude Code, Devin or GitHub Copilot directly in group chats to ship fixes. It is on all Slack plans, requires human approval for production merges, and archives the channel once work is approved (Gizmodo).
Grok Bot left beta and became a bundled feature, now included with SuperGrok Plus, SuperGrok Heavy, Cursor Pro+, Cursor Ultra and Cursor Teams plans (xAI).
The maintainers of IPFS at Interplanetary Shipyard are winding down their work on the protocol (Interplanetary Shipyard).
