The company that stopped waiting for a better coding agent

⚡ 60-Second Catch-Up

A fintech company built its own coding agent and is now ahead of the labs that sell them. Ramp's internal harness, Inspect, raised 75% of all merged pull requests at the company as of May, up from about 60% in January, crossed a million sessions in July, and is maintained by five and a half people — four engineers, a director and a part-time PM — while more than 150 Ramp engineers have contributed to it. Roughly 80% of Inspect is now written in Inspect. Gergely Orosz, who reported the story with Jessica Salmon and Ivan Klaric, gets the founding team to explain why they turned down Claude Code, Cursor and Codex: not model quality, but that local machines cap you at one or two parallel sessions, and that a third-party harness cannot see your telemetry, your feature flags or your read-only prod replica. So Inspect runs on remote sandboxes that spin up in under five seconds and verifies its own work — running tests, querying flags, screenshotting frontend changes. → The transferable part is not "build your own harness." It is the observation that the ceiling on your agents is access, not intelligence. Before you pay for a smarter model, write down the three things you check by hand after every agent run — the test suite, the staging screenshot, the log line — and ask whether the agent can reach any of them. That list is your actual roadmap.

Generation speed became a product category this week, and the cheap tier got very loud. NVIDIA put Groq 3 LPX into full production, an inference accelerator aimed squarely at the decode phase — how fast tokens come out for one user — hitting a record 3,400 output tokens per second on Gemma 4 31B at a 100,000-token context, which NVIDIA says is four times the nearest alternative. Nebius is first to deploy. At the other end of the market, an anonymous model called Ox Alpha ran 26 trillion tokens through OpenCode in four days across 327,000 users, free, from a maker nobody has named. And Tomasz Tunguz argues the whole stack is caught in an AI bullwhip — GPU scarcity cascading into memory, storage and datacenter construction costs. → Latency is about to stop being a fixed cost you design around. If your product has a step where the user waits on a model — a summary, a search, an agent turn — you have been building the UX for a constraint that is being engineered away. Time that step today and write the number down, because the version of your product that assumes it is three times faster is a different product.

Everyone is discovering that agents need documentation written for machines, not for people. A design team rebuilt its design system into layered Markdown files so agents could reason over it, after finding that AI-generated frontend code looked right but kept missing hover states, spacing tokens and button rules. Ron Bronson argues agent systems need failure affordances — an explicit channel for "something went wrong", separate from the permission layer that decides what an agent may do — citing an agent that discovered it could cancel other users' gym reservations. And the skills themselves are going multilingual: the share of newly created agent skills written in a language other than English rose from 13.0% in the first quarter of 2026 to 16.3% in the second, with Chinese leading. → Your README is now an API surface. Pick the one internal doc your agents get wrong most often — the deploy runbook, the component rules, the data model — and rewrite it as explicit constraints rather than prose explanation. "Buttons use --space-2" beats three paragraphs on spacing philosophy, and it is the difference between an agent that is almost right and one that is right.

🛠 TOOLS — 11

AI & agents

• ProtoNote ⭐ — Share an AI-built prototype or artifact with your team and collect pinned notes directly on the page. Drop in an HTML file, screenshot, Markdown or PDF and it becomes a reviewable link — or skip the download entirely and tell Claude to make one, since it runs as a Claude connector. Links are private by default and visible only to the people you send them to · Free
• Rome — Runs persistent AI agents, workflows and apps inside a guardrailed collaboration environment — an agentic OS rather than a single-session harness, aimed at agents that stay resident instead of dying with the terminal window · Open source (MIT)
• Kern ⭐ — A rootless container runtime with kernel-enforced isolation and cold starts around 3.5 ms, with no daemon. Built for running untrusted workloads — which, as of this year, includes whatever your agent just decided to execute · Open source (Apache-2.0)

Build & ship

• PicoMQ — Durable, real-time streaming over plain HTTP backed by S3-compatible object storage, with a CLI for node management and stream operations and a Docker path in. Durable streams without standing up Kafka · Open source (Apache-2.0)
• LatticeDB — An embedded single-file knowledge-graph database with vector search and full-text search built in — roughly what SQLite is to relational data, pointed at graphs · Open source (MIT)
• Maiao — Brings a Gerrit-style review workflow to GitHub, GitLab and Gitea from the command line, so stacked changes stop being a branch-management exercise · Open source (MIT)
• fdeploy — Self-hosted deployment automation for Windows and IIS — a genuinely underserved corner of the deploy-tooling world, where most options assume Linux and a container registry · Free (self-hosted)

Design & create

• Appshot Gallery — A curated gallery of App Store screenshots, filterable by visual style — 3D, pixel art, hand-drawn, brutalist, dark mode — for when you are designing store listings and need to see what converts in your category before inventing something · Free
• Acebuilder — Describe a landing page and get it built from Aceternity UI Pro templates as production-grade React and Tailwind, then reshape layout and copy in chat. Starts from templates and component blocks rather than a blank page, and exports the code · Paid (credit-based; heavier models and longer sessions cost more credits)
• TeXbrain — A browser-based LaTeX editor running pdfTeX in WebAssembly, with live preview, real-time collaboration and git integration. No TeX install, no server round-trip · Open source (MIT)

Growth & ops

• GlassBox — Runs every measurement a website can take from your browser and shows the results back to you — canvas, audio, codec and math fingerprints, hardware signals, session data — then estimates how identifiable you are. The same class of signal anti-fraud and tracking scripts collect, surfaced rather than hidden. Entirely client-side · Free

📚 RESOURCES — 8

Steal the template

• How To Report A Bug So It Actually Gets Fixed — a structured process for bug reports: own the bug, build a reproducible environment, bisect the version, document thoroughly, and assume the problem is yours until proven otherwise. Worth handing to anyone who files issues against your project, and worth following when you file them against someone else's.
• How to Make Your Design System Agent-Ready — the actual restructuring: splitting design system documentation into layered Markdown files so an agent can reason over the rules instead of pattern-matching to "almost right". The failure cases named — hover states, spacing tokens, button variants — are the ones every Figma-to-code pipeline hits.
• How to Choose a UI Font (and 10 Inter Alternatives) — two concrete criteria you can check rather than eyeball: cap-centred vertical metrics, so labels optically centre on buttons and align with icons off a single padding token; and an x-height at 70–75% of cap height, so words stay balanced when centred on capitals. Then ten fonts that meet them.

Learn the craft

• Reinforcement Learning for LLMs: The Complete Guide — Cameron Wolfe's 98-minute survey of RL as it is actually used on language models: the foundations, the algorithms including policy gradient methods, and how alignment, reasoning and safety work got folded into training. Long, and the closest thing to a single reference on a field that moves monthly.
• Speculative Programmatic Tool Calling — pre-launching tool calls during token generation so non-blocking calls run in parallel, JIT-compiler style, for a 1–1.2x runtime speed-up. Most useful on memory-bound local models and high-volume tool loops, and a clear write-up of where agent latency actually goes.
• How to Design an Animation — breaking intricate motion into components you can express mathematically, so interface animation becomes something you reason about rather than tune by feel.
• Why Code Verification Matters More Than Ever in the Age of AI — the counterpart to every "AI writes most of our code now" statistic: what verification has to look like when generation stops being the constraint.

Benchmarks you can re-run

• Graph Engineering — a curated collection of papers, benchmarks and open-source projects on using dynamic graph structures to organise tasks, coordinate agents, track runtime state and evolve multi-agent systems. A reading list and a benchmark set in one repo.

📖 READS — 22

☕ Under 5 minutes

• The truth is we are in sales — tie creative work to outcomes defined before the work begins — revenue, funding, engagement, stakeholder buy-in, approvals — rather than to subjective judgments of quality afterwards. Uncomfortable and correct.
• The AI Bullwhip — Tomasz Tunguz traces how GPU scarcity propagated into server shipments, memory allocation and datacenter construction costs, with each link amplifying the last. If you are forecasting infrastructure spend, this is the shape of the curve.
• Tooltips need a delay, and then they need to skip it — a small interaction problem worked all the way through, and a good example of the category of detail that separates interfaces that feel considered from ones that do not.
• Myako Display Font Specimen and Bold Typography — oversized curves and deep inktraps, built for headlines, packaging and branding that need to hold up at size.

🍵 5–10 minutes

• LLMs could control their host machines by exploiting inference engines — machines running frontier models are unusually valuable targets: enough compute to serve a model, direct access to the weights, and privileged network position inside the datacenter. The research shows a model can emit token sequences that exploit the loader software itself. A genuinely new attack surface.
• Agent Experience Needs Failure Affordances — an exception channel is not the same thing as a permission router. One decides whether an action is allowed before it happens; the other lets the agent say something has gone wrong afterwards. Most systems have built the first and skipped the second.
• Anger, Anxiety and Agency — Armin Ronacher on the emotional weather of building right now, from someone who has been shipping open source long enough to have perspective on it.
• Hot Chips 2026: CUDA Targets RISC-V — Nvidia extending CUDA to RISC-V would let RISC-V CPUs feed GPU compute, though the honest read is that most existing RISC-V hardware will not meet the requirements and the software ecosystem is still behind x86-64 and aarch64.
• Being a Senior Designer Doesn't Make You Safe Anymore — fifteen years of specialised skill was a moat until the specialised part became widely accessible. The argument about sunk cost and status quo bias applies well beyond design.
• How Europe is killing makers and micro-entrepreneurs — new EU packaging regulations impose costs and administrative load that scale badly downward, to the point where some micro-businesses are considering selling only outside the EU. Regulatory cost as a founder problem.
• When str.lower() is a security vulnerability in Python — Seth Larson on Unicode case folding turning a routine normalisation call into an exploitable one.
• C2PA Cameras Do Not Survive Contact with Reality — David Buchanan on content provenance signing, and what happens to the trust model when the signing hardware is a phone you control.
• An unlikely experiment — short, personal, and worth the six minutes.
• We used to log off — the internet stopped being a destination with an entrance and an exit. Infinite scroll, persistent notifications and seamless everything each made things more convenient and collectively removed the edges.

📚 Longer, worth it

• Why Ramp built its own in-house coding agent, Inspect — the full account behind today's lead: the pivot from a Chrome extension for designers to a remote development environment with an agent on top, the Cloudflare Durable Objects and Modal sandbox architecture, the tricks that get a fully provisioned environment up in under five seconds, and a collaboration model where every session is public and there is no opt-out.
• When code is abundant — GitLab on what changes when generating code is cheap: the hard part moves to trusting and verifying it, and the organisations furthest along — Stripe, Spotify, Amplitude are named — are the ones that built governance and context systems rather than just adopting tools.
• The Economics of the Intelligence Frontier — Sean Linehan's argument that once a model passes the intelligence a task actually requires, that task commoditises and competition moves to cost, latency, infrastructure and distribution. Frontier labs can still be enormous businesses, but only if new capability opens valuable markets faster than rivals reproduce the old ones.
• AI Coding will Prevent Expertise — the uncomfortable loop: using AI coding tools well appears to require expertise that was previously built by doing the work the tools now do.
• Your executable is a SQLite database and Queryable Executables — Farid Zakaria replaces ELF with SQLite as an executable format in a prototype called SELF, which makes binaries something you can query and modify with SQL. Two posts, consecutive days, and the second is where it gets real.
• What language are agent skills written in? — the measurement behind today's 16.3%, and a useful early read on where agent tooling contributions are coming from.
• Stop Using Conventional Commits — the argument that scope matters more than type, and that putting type first produces messages that read as structured while communicating less than a plain sentence would.
• OpenAI Jalapeño: Better than Nvidia Blackwell — SemiAnalysis on OpenAI's custom silicon, and a useful counterweight to reading only Nvidia's own framing of the inference market.

📊 NUMBERS — 11

• 75% — share of merged pull requests at Ramp raised by Inspect, its in-house coding agent, as of May, up from about 60% in January, two months after the agent's v2 launch (Gergely Orosz, reporting with Jessica Salmon and Ivan Klaric).
• 5.5 people — the team maintaining Inspect: four engineers, a director and a part-time PM. More than 150 Ramp engineers have contributed to the codebase, and over 80% of Inspect is written using Inspect (Gergely Orosz).
• 1,000,000 — Inspect sessions passed in July, with more than 200 internal agents now built on the platform (Gergely Orosz).
• Under 5 seconds — time to spin up a fully provisioned remote development environment, including Postgres, Redis, RabbitMQ, Temporal, Chromium and VS Code Server (Gergely Orosz).
• 3,400 tokens per second — record output speed for Gemma 4 31B at a 100,000-token context, measured by Artificial Analysis on NVIDIA Groq 3 LPX; NVIDIA states this is four times the nearest alternative platform for latency-sensitive workloads (NVIDIA).
• 500 MB — SRAM packed directly onto each Groq 3 chip die to sidestep memory bandwidth limits, with 256 chips per LPX rack, fabricated by Samsung (CNBC).
• 26 trillion — tokens processed through the anonymous Ox Alpha model on OpenCode in its first four days, across 327,000 unique users and 8,328,244 completed sessions. That is second place on OpenCode, behind DeepSeek V4 Flash at 33 trillion. On OpenRouter, a separate 11.6 trillion tokens over three days made it the largest launch in that platform's history, against 4.4 trillion for the next biggest (OpenCode, OpenRouter).
• 16.3% — share of newly written agent skills in a language other than English in Q2 2026, up from 13.0% in Q1, measured across 255,068 skills. Chinese leads the non-English share. For scale, GitHub-wide non-English documentation took ten years to go from 3.7% to 13.0% (Plicara Labs).
• 3 cents — price per email sent on the outbound platform Explee, against a seat-based benchmark of 351 sales teams posting 60% quota attainment and 40% annual attrition (Dru Riley).
• 39% — share of key job skills the World Economic Forum expects to change by 2030 (World Economic Forum).
• $1,000,000 — size of Goodfire's new research grant program for AI interpretability, which also includes free access to its Silico research platform (Goodfire).

📰 WHAT HAPPENED — 14

Models & math

• NVIDIA's Groq 3 LPX inference accelerator entered full production, extending the Vera Rubin NVL72 platform and targeting the decode phase of inference specifically. Nebius is the first AI cloud to adopt it, with racks going live later this year (NVIDIA).
• An anonymous model called Ox Alpha became the largest launch in OpenRouter's history, generating 11.6 trillion tokens there in three days, while separately running 26 trillion tokens through OpenCode over four days — free, via an OpenAI-compatible endpoint. Its maker, release date and provenance are all still unlisted (OpenCode, OpenRouter).
• Alibaba launched Wan3.0, a video model that generates 30-second clips from text, following a record $10 billion share sale (Yahoo Finance).
• Anthropic hired Amir Salek, the engineer who founded Google's custom-chip program and ran its TPU business, for its own silicon effort (RuntimeWire).
• Nvidia detailed plans to extend CUDA to RISC-V at Hot Chips 2026, opening a path for RISC-V CPUs to feed GPU compute (Chips and Cheese).

Safety & policy

• Goodfire opened a $1M research grant program for AI interpretability, bundling free access to its Silico platform (Goodfire).
• New EU packaging regulations landed on makers and micro-entrepreneurs, with compliance costs that scale badly for very small sellers — enough that some are weighing selling only outside the EU (Lectronz).
• Nitter and XCancel received cease and desist notices, putting the main privacy-preserving front-ends for X at risk (Nitter issue tracker).

Business moved

• Ramp disclosed that 75% of its merged pull requests come from its own internal coding agent, as of May, built by a 5.5-person team rather than bought from a frontier lab (Gergely Orosz).
• The FDA authorised the first wearable that continuously monitors both ketone levels and blood sugar, a first for combined continuous metabolic monitoring (FDA).

Platforms shifted

• Apple introduced the M6 and M5 Ultra, alongside a new Mac Studio with M5 Max and M5 Ultra and a Mac mini with M6 and M5 Pro. The M6 is Apple's first Mac chip on a 2nm process, and a refreshed iMac with the same chip is expected before the end of the year (Apple).
• Adobe made Firefly's Generate Music, Generate Speech and Generate Sound Effects generally available, building instrumental tracks from descriptions or footage and converting scripts to voiceover in more than 20 languages, inside one workspace (Digital Trends).
• Figma added two auto-spacing options to auto layout, "Around" and "Evenly", alongside the renamed default "Between" — matching distribution patterns CSS already supported (Figma).
• Firefox 157 will ship JPEG XL enabled by default on all platforms, ending a long standoff over the format (Mozilla dev-platform).