The Collection

Gergely Orosz reprints OpenAI’s Asana case and then goes looking for the human who lived it. Codex helped Asana replace Enzyme with React Testing Library in two weeks. Model and infrastructure costs about $12K. The previous staffing estimate: roughly $6M across something like five years. Orosz’s first reaction is the face you make at four engineers times five years for a test migration. His second reaction is better. Talk to Dan Ubilla, who leads developer productivity at Asana.

Asana started in 2024. Four thousand-plus Enzyme files. A first LLM pass got about a quarter moved, plus better mocks and coverage. The remainder dropped to opportunistic: nice to have, behind critical and important migrations. Two years in, five years was a rational guess for when that shelf might clear if the priority never rose. The $6M was back-of-envelope: time per file, times remaining files, times an hourly rate. It assumed almost no learning curve and no models. It was a baseline for work nobody wanted to own.

Opportunistic is how a company says this will never start.

Airbnb, last year, moved 3,500 Enzyme files in six weeks with LLMs. Hand estimate: 1.5 engineering years. Uber moved 600,000 JUnit tests, fifteen million lines, in four months with two engineers and AI. Bun’s Zig-to-Rust rewrite: 530,000 lines in two weeks for a $165K API bill. Orosz’s point is not that OpenAI’s arithmetic is sacred. It is that pre-AI, these migrations were impractical. Post-AI, the opportunistic shelf is where you go to prove the tool, then where you go to clear the real backlog. Engineers still design the loops. Verification still sits with people. The years-long window where you support two libraries at once just got shorter. That alone can be worth the bill.

A coding agent is rarely told where the change belongs. It searches for a name, opens files, guesses the graph, searches again. Grep matches characters. That is all it does. The failure modes are three: a flood of irrelevant hits, a location that shares no text with the query, and a match on the wrong symbol. The first and third burn tokens. The second ships a bug that still compiles.

Sonar Vortex treats the codebase as a graph: classes, methods, interfaces, calls, references, exact file and line. Built without a compiler, so it stays usable on mid-edit code. A controlled comparison used real merged commits as ground truth, prompts without file names, ten runs per side, build and tests required to count. Across six tasks in four languages, cost fell in every case: a Java interface change 36% cheaper, a package rename 20%, Python 20%, TypeScript 5%, a Java argument-order fix 15% on the typical run, C# 20% on the typical run. Where finding code was not the bottleneck, cost stayed within a few percent either way.

Passing tests is not the same as finding every site that needed to change.

The wins clustered where a change had to land identically across every implementor of an interface. Text search cannot cleanly list those. The open question Kapur leaves on the table is not speed. It is how you would know, concretely, whether the agent found everything. A developer sees reread files. A leader sees costs that wander. A product manager sees a defect with no obvious origin. Same root cause, three seats.

Token pricing began at the model layer. OpenAI’s API in 2020 metered raw inference. ChatGPT later sparked applications that combine data, tools, orchestration, and workflow. When those products still bill in tokens, a16z argues, they train the customer to compare the application with raw compute. They give away the value of data and workflow. They expose buyers to forecasts they cannot make. Margins stay weak while model costs fall.

Their rule: price at the highest layer of value you can reliably measure, attribute, and defend. Sell model access, price tokens. Turn models into useful work, price a recognizable unit, often through credits. Deliver a clear business result, price the outcome. In a survey of 50 technical AI buyers, 27 preferred credits tied to recognizable work. Fourteen preferred tokens. A support lead can estimate conversations. Context length, retries, and reasoning tokens are a different sport.

You are not a model. Do not price per token.

Credits only help if they map to work the customer already understands. A weak credit is cost-plus tokens with a new name. Clay’s 2026 memo, as a16z recounts it, separated Data Credits from Actions after years of mispricing the Pro segment. Fixed pricing where costs are predictable. Token pass-through, without markup, for volatile reasoning models. Multiple meters are fine. The wrong meter for the wrong layer is the failure. Transparency does not require the billing meter and the cost meter to be the same thing.

Eric Newcomer has two personal assistants in the same fever. Town, the enterprise personal assistant from former Plaid CTO Jean-Denis Greze and former Google applied AI product director Tony Vincent, is in talks to raise at a $1 billion valuation in a round led by Index, sources tell Newcomer. Earlier money from Andreessen Horowitz and Forerunner. Instinct, the text-based consumer assistant from Spear Street Technology and former Sierra researcher Noah Shinn, has cumulatively raised $350 million at a $2.5 billion valuation, as the Wall Street Journal first reported. Index and Benchmark lead the latest round.

Newcomer is surprised Index is in both. An investor shrugged: Greylock once held Facebook and LinkedIn. Right now the power users look like the venture capitalists themselves. The bet is generation-defining applications. We print the sourcing as sourcing. In talks is not closed. Sources tell us is not a term sheet on the desk.

The hottest product in the room is still whoever answers first.

Thursday we asked who the agent serves when the developer’s interest diverges. Today the money is simply louder. Two products. One category. A billion and two and a half billion before most people have a clerk they trust with the calendar.

The pilot works in the demo. Then it meets the real process. A different spreadsheet. An approval that still lives in email. Exceptions that go to the employee who has handled them for years. AI Adopters opens a six-part series for CEOs with a blunt test: if the person who knows how this process works were unavailable for two weeks, could the rest of the team still run it. If the answer is no, you are not ready to automate. You are about to automate one person’s memory and hope they never take leave.

Deloitte’s August 2026 survey of 501 U.S. leaders at organizations already piloting agents: only 5% said their business processes were highly prepared. Just 21% called them prepared or highly prepared. PepsiCo, McDonald’s, and Stripe are the named cases above the fold. PepsiCo moved expensive engineering decisions into simulation before changing a physical facility. McDonald’s put voice automation into a live drive-through and ended an IBM test without publishing final operating results. Stripe broke compliance reviews into small questions, let AI gather evidence, and kept people responsible for every answer.

Five percent highly prepared. The cases sit behind the wall.

Below the paywall: the decision PepsiCo moved before committing physical capital, why McDonald’s early 85% accuracy figure does not describe its final test, how Stripe redesigned a white-collar review without automating the final judgment, and a CEO approval test. We have the free frame and the Deloitte numbers. We do not invent the rest.