The Collection

The AI Corner’s letter starts with a bug. Ramp’s coding agent, Inspect, wrote 75 percent of the code changes Ramp accepted by August 2026, according to the letter. An engineer asked it why a cash-planning feature kept running twice. To answer, the agent opened eight working sessions across three codebases, used about 5.6 million tokens, cost $93, and shipped two approved fixes. The session the engineer started cost about a fifth of that. Both fixes came from sessions the agent opened on its own.

Ramp still could not say whether $93 was a good deal. If the bug was costing customers money every day, it was cheap. If nobody would have noticed for months, it is hard to call a bargain. So Ramp built the record first. It stitches linked sessions into one complete run, then has a model split each run into work items: one goal, one result, an owner and a share of the cost. Over three weeks it did this for about 200,000 runs. Only 17 percent of runs pursued more than one goal, and those runs took 58 percent of the spend. Ramp has not published an ROI number. What it built is the ledger you would need to calculate one.

Measure per goal, not per session.

The letter is hard on the usual dials. Tokens measure consumption. Tasks completed mean little if a third of the tickets reopen. Time saved is the worst: it cites METR’s 2025 study, in which 16 experienced programmers took 19 percent longer with AI tools and afterwards believed they had been 20 percent faster. It cites Klarna, whose chief executive said in 2025 that chasing cost had lowered quality. Its test is one sentence you could defend to a board: we spent X for Y, accepted results at Z each, against W the old way, and the metric moved by N. The letter is sponsored by a CRM vendor; the stand skips the pitch.

Semafor shows where the habit has landed. A Selby Jennings survey of 500 companies found a quarter of US financial firms have tied raises, promotions and bonuses to AI proficiency, and nearly a third have created roles or departments to track AI use. Semafor’s phrase for the result: managers as part-time AI accountants. Tokenmaxxing is dead, an IBM post declared in June. The accounting has not caught up.

Trilogy AI ran a demanding migration and wrote down what each model actually shipped. The job: move the Engineering Behavior Observatory, which runs coding agents and records their work, onto Harbor, a framework that evaluates agents in isolated environments. A pass meant running all six existing agent integrations inside those environments and letting an independent verifier decide success. The study combined 21 attempts across Meta’s Muse Wonder, Google’s Gemini 4 Argon and Anthropic’s Claude Opus 5.5. The author flags the limits: Opus had more time, and the routes and context settings differed.

The finding that matters is about Muse. Its submissions through Pi showed hundreds of passing repository tests. Both lacked the production connection to the real coding-agent integrations. The tests used a scripted executor that edited files and generated synthetic records, so they passed while the code needed to run the real agents was absent. One Muse run on DeepSeek’s harness did the same. Argon built more, but ran the agents on the host while Harbor verified inside a container, and one defect let re-evaluation report success without calling the verifier at all.

A model can write the tests, then pass them around the hardest requirement.

The cost line reads like a receipt. Reviewed Muse runs cost roughly $3 to $8 and ran 10 to 38 minutes. Native Argon runs cost about $12 each over roughly 70 minutes to two hours. Opus took between about two hours twenty and just under three hours, at an equivalent metered cost of roughly $42 to $57, and delivered the working core of the migration, with defects still to repair: two submissions could treat an empty result as success. The comparison excludes review and repair work, which is exactly the cost the cheap runs push downstream.

Cache reuse swung from about 9 percent to about 99 percent depending on which harness carried which model. An advertised token price, the author notes, gives an incomplete picture of cost in a coding harness. Put that beside Ramp’s ledger and the lesson is the same from both ends. The cheapest run is the one whose output you can accept, not the one with the smallest bill.

a16z’s letter calls the long customer journey a relay race: weeks or months of handoffs, each one a place for value to leak. Its worked example is plaintiff law. Leads go to the firm that calls back first. Paralegals give up on the fifth unanswered call to a witness. A routine claim settles on the first offer because nobody had time to read the doctor’s note that hinted at a brain injury. Bills drop out of the demand letter.

The pitch is a single long-running agent assigned to the case from the intake call, with the goal of maximising the client’s recovery. It saves and retrieves context, wakes on events and deadlines, sends sub-agents to place the tenth or twentieth call, drafts documents for review, and keeps the whole record. The commercial turn is the point of the letter: when the agent drives the outcome, the vendor can price on outcomes and claim a share of revenue, not just of labour savings.

Pay on the outcome, and the receipt finally has a total.

The letter is candid about the friction. Most US states bar non-lawyers from sharing legal fees, so pricing would have to track settlement values without taking a cut. Procurement teams lack precedents for outcome contracts, and attribution fights will come. Its historical case is Red Ventures, which owned whole insurance sales journeys and was paid per policy closed: about $370 million EBITDA on about $1.4 billion in sales in 2022, per S&P Global, falling to an estimated $160 million in 2025. a16z reads that as a model under threat from agents. It is an investor’s thesis, and the letter says it invites pitches.

Read against The Receipt, the relay is the one place where the sentence for the board writes itself. A settlement is a number. The open question is who checks the work before it reaches the finish line.

Gabriel Unger, an economist at Stanford’s Digital Economy Lab, and Arvind Narayanan wrote the longest letter in the bag and the sharpest answer to the relay. Their charge: the dominant program treats AI as almost entirely about automating human labour, and much of the industry’s weak return comes from that bias. They state a trilemma for agents sold under it: completely autonomous, reliable in high-stakes settings, truly general purpose. Pick two.

The phrase worth keeping is about the loop. Under delegation, they write, the human is technically in the loop, but just standing at the starting line and then waiting at the finish line. Their alternative is a research agenda with four problems: white-box agents that show their sources, assumptions and uncertainties; agents fast enough for real conversation, so people keep their flow; continuous steering beyond turn-taking, with shared workspaces and interruptions; and benchmarks that score the human and the model as a team.

If the human only signs at the finish, the receipt is all they ever see.

White-box agents need process traces: records of how the work is done, not only what it produced. Coding had public repositories full of them. Law, medicine and most sciences do not. Semafor’s interview with Reflection AI’s Ioannis Antonoglou lands on the same gap from the business side: companies capture work products, not the work itself, the calls and shared context that link one action to the next. He expects the rebuild to take a decade or two.

The authors are not against automation where the work is routine and easy to verify. They name Thinking Machines as a lab with a published architecture for collaboration agents and no shipped product yet. The stand takes one line from them and adds one of its own. Regulation that requires human oversight, they write, must insist on effective oversight instead of rubber stamping. A tick at the end of a relay is a stamp. A hand on the wheel is oversight.