MECHANIZE

Institutional Archetype

THE ENVIRONMENT SUPPLIER — Most of the evaluators on this map grade models. Mechanize builds what models are trained and graded on: tasks with a prompt, a sandbox and a grader, bundled into environments that frontier labs use “either to train their models directly or to measure their capabilities.” It was founded in April 2025 by three people who had come from Epoch AI, the nonprofit that builds benchmarks, and it sells the same two things, environments and evals, to the party Epoch measures. The file is open on the seam, not on a fault: the supply side of evaluation is now a venture-funded company with named individual backers, a published goal of automating work, and a September 2026 press report that Google completed a talent deal for it.

Mandate & Origin

Mechanize, Inc. announced itself on April 17, 2025, signed by Matthew Barnett, Tamay Besiroglu and Ege Erdil, as “a startup focused on developing virtual work environments, benchmarks, and training data that will enable the full automation of the economy.” Its stated method: “creating simulated environments and evaluations that capture the full scope of what people do at their jobs.” Its stated market: “workers in the US are paid around $18 trillion per year in aggregate” (Mechanize, Announcing Mechanize, Inc.). Tamay Besiroglu described it the same day in his own post as building “virtual work environments, benchmarks, and training data to enable the full automation of all work” (Besiroglu on X), and in June 2025 wrote that the company’s goal is “to automate all work as quickly as possible” (Besiroglu on X). Epoch’s own page on its co-founder says he “left Epoch to co-lead Mechanize, a startup building virtual work environments, benchmarks, and training data for AI development” (Epoch AI). Ege Erdil is a signatory; see his profile for his Epoch history.

The home page now describes a narrower product: “We build environments and evals for frontier coding agents. In these environments, models carry out software engineering work such as building a feature, deploying an application, or debugging an issue in an unfamiliar codebase. A grader scores the model’s performance, and these scores serve as signals during reinforcement learning and evaluations” (Mechanize, home). Its one named public product is GBA Eval, “a benchmark that measures how well coding agents can write a Game Boy Advance emulator from scratch in 24 hours” (Mechanize, home); the benchmark’s own site says “Their emulators are graded against Mesen2” (GBA Eval). The company’s pages list a podcast episode of June 16, 2026 and eight essays dated May 30 to November 25, 2025 (Mechanize, podcast; Mechanize, essays).

Staff and work. “Mechanize is a team of around 35 people including engineers and other staff.” Each task is “owned by a single engineer”; “We think of 40 hours of in-person work as a reasonable baseline for a work week”; “most of your time will be spent prompting coding agents” (Mechanize, working here). Its open roles are in San Francisco and in person: a software engineer ladder from $300K to $400K salary, “Top performers can earn more from bonuses than salary,” a $200K recruiter, and a Research Engineer, Alignment at $300K to $400K who will “create evals that test for misaligned model behaviors,” such as “acting sycophantic, being dishonest, or taking unsafe actions” (Mechanize, apply; Mechanize, software engineer; Mechanize, research engineer). The interview is “3 stages, consisting of 5.5 hours in total” (Mechanize, interview process). The contact page lists a general inbox and a “Direct contact for Guive Assadi” and gives no titles (Mechanize, contact).

Funding & Backers

• Round. “We’ve raised $9.1 million at a $500 million post-money valuation from Marco Mascorro, Adam D’Angelo, and Devendra Chaplot,” signed Matthew Barnett, Tamay Besiroglu and Ege Erdil, April 24, 2026 (Mechanize, press release).
• Backer list. The April 2025 announcement: “Mechanize is backed by investments from Nat Friedman and Daniel Gross, Patrick Collison, Dwarkesh Patel, Sholto Douglas, and Marcus Abramovitch” (Mechanize, Announcing Mechanize, Inc.). The home page now: “Mechanize is backed by Nat Friedman and Daniel Gross, Patrick Collison, Adam D’Angelo, Marco Mascorro, Dwarkesh Patel, Sholto Douglas, Devendra Chaplot, Alex Atallah, and Marcus Abramovitch” (Mechanize, home).
• Revenue and customers. None of Mechanize’s pages states revenue, a customer or a contract. The only description of the buyer is “frontier AI labs” (Mechanize, working here).
• The reported deal. Business Insider, August 5, 2026: “Google is in talks for a $1.5 billion-plus deal” and a “potential deal that would involve Google hiring some of Mechanize’s talent” (Business Insider, Aug 5). September 11, 2026: “The final terms of the deal weren’t disclosed” (Business Insider, Sep 11).

Institutional Voice & Intent

Mechanize on itself and on what it opposes. Its October 6, 2025 essay: “Autonomous agents that fully substitute for human labor will inevitably be created because they will provide immense utility that mere AI tools cannot. The only real choice is whether to hasten this technological revolution ourselves, or to wait for others to initiate it in our absence.” Its image for policy: “humanity is more like a roaring stream flowing into a valley, following the path of least resistance,” and its limit on regulation: “Delaying or regulating its use may be possible, but forgoing the technology entirely seems to be beyond our capabilities.” It closes: “Little can stop the inexorable march towards the full automation of the economy. We should be glad,” then “Want to help accelerate the inevitable?” with a hiring link (Mechanize, The future of AI is already written). On its product thesis, from July 2025: “Teaching AIs these new capabilities will require the dedicated efforts of high-skill specialists working full-time, not low-skill contractors working at scale,” under the heading “Good RL environments are the bottleneck” (Mechanize, Sweatshop data is over). On its critics, from November 2025: “Our critics tell us that our work will destroy the world,” and of the book If Anyone Builds It, Everyone Dies, “if their theory about deep learning only makes predictions about future superintelligent AIs, with no testable predictions about earlier systems, then it is functionally unfalsifiable” (Mechanize, Unfalsifiable stories of doom). On its own quality bar: “being rigorous about this step is what distinguishes our tasks from public benchmarks scraped from open-source repositories,” and “we’ll increasingly automate parts of the task creation process itself using AI” (Mechanize, working here).

Stated intent: “the full automation of the economy,” by selling the environments and evals on which agents learn and are scored, and to hire “to help accelerate the inevitable.”

Observed intent: Be the supplier of the training and scoring tasks for frontier coding agents. By its own pages it employs about 35 people, hires in person at $300K to $400K, bundles tasks into environments labs use “to train their models directly or to measure their capabilities,” and posts a benchmark of its own. It also hires evaluators of misaligned behavior, so the same company is a source of capability signal and a tester of its failure modes.

Gap: The stated goal and the product line up: the essays argue for full automation and the company sells the inputs to it. The structural point is that its customer list is undisclosed while its category sits on both sides of evaluation: a task is used to train a model and, in the company’s words, to “measure their capabilities.” Epoch’s FrontierMath record on this site is the closest earlier case of a test whose maker and buyer overlapped. The record here shows no such overlap by name: no customer is on any Mechanize page. The second gap is in the record, not in any statement: a first-tier outlet reports a completed talent deal with Google, and the company’s press page ends in April 2026 with no mention of it. Neither source says what the arrangement covers.

Interest: Mechanize states positions in its essays; each is set out below from its own pages, with what the position pays, documented, and then its case in its authors’ words.

• Full automation is inevitable and should be accelerated. Position: “Autonomous agents that fully substitute for human labor will inevitably be created,” and “We should be glad” (Mechanize, The future of AI is already written). What it pays: the company’s product is the environments and evals that train agents on jobs, so a rule that leaves labor-substituting agents unrestricted keeps its category open, and the same essay carries the company’s hiring link. Best case, theirs: “The only real choice is whether to hasten this technological revolution ourselves, or to wait for others to initiate it in our absence.” Who else gains: the labs that sell or deploy agents, which Mechanize says are its buyers (Mechanize, working here).
• Training data should come from full-time specialists building software environments, not contractors. Position: the July 2025 essay above. What it pays: Mechanize pays salaried engineers to build tasks, one engineer per task (Mechanize, working here; Mechanize, apply), so a buyer norm that favors that model favors the company’s own staffing. Best case, theirs: “Software, not datasets,” because interactive environments “remain valuable as models improve.” Who else gains: other vendors that build environments with salaried specialists; the page names none.
• The AI-doom argument is not evidence. Position: the November 2025 essay above. What it pays: the company sells capability inputs to labs, so an argument against the case for restraint is an argument for the market it sells into; the same company lists a Research Engineer, Alignment role to “create evals that test for misaligned model behaviors” (Mechanize, research engineer). Best case, theirs: “We want to engage with these critics.” Who else gains: any lab or investor whose plans a restraint rule would delay. The essay does not mention one.

Interest is a documented fact about each position, not a claim about motive, and not a verdict that the argument is wrong. Each best case is stated at the strength its author states it.

Position in the Apparatus

• The supplier beside the grader. The evaluators on the evaluators map grade models. Mechanize builds the tasks and graders that models are trained on and scored with, and says so: the grader’s scores “serve as signals during reinforcement learning and evaluations” (Mechanize, home). It is the only entry whose product is both the thing a lab trains on and, on its own account, the thing a lab measures with.
• The Epoch lineage. The company was founded by Barnett, Besiroglu and Erdil. Epoch’s page says Besiroglu “left Epoch to co-lead Mechanize” (Epoch AI); Erdil’s Epoch history is on his profile. The nonprofit and the company share a vocabulary (environments, benchmarks, evals) and, per the nonprofit’s own transparency page (Epoch AI, transparency), a client class: Epoch AI lists OpenAI, Google DeepMind and xAI among its commissioning parties. The two organizations are separate; no shared funding or contract is documented.
• Other evaluators it sits beside. Andon Labs runs agents in real businesses and has three labs citing its tests in release documents; METR sets autonomy measures. Mechanize names no lab as a customer and appears in no lab release document located.
• The Google report, as the sources state it. Business Insider, September 11, 2026: “Mechanize’s co-founder and ex-CEO, Tamay Besiroglu, now works as a research scientist at Google’s AI division, DeepMind, according to his LinkedIn profile. Over a dozen other former Mechanize employees now work at Google, per LinkedIn, mostly on the company’s ‘midtraining’ efforts” and “Google and Besiroglu declined to comment.” It adds: “Mechanize’s former chief of staff, Guive Assadi, now describes himself as its CEO, per LinkedIn” (Business Insider, Sep 11). Besiroglu’s own account is on his profile. The article says Mechanize’s technology helps tech companies’ AI models improve at coding, “a challenge Google has struggled with” (Business Insider, Sep 11). Mechanize’s own press page lists no announcement after April 24, 2026 (Mechanize, press releases), its press page lists four items of outside coverage and no Google item (Mechanize, press), and its contact page names Assadi without a title (Mechanize, contact).

Actions & Leadership Choices

CONDUCT: DEAL REPORTED BY THE PRESS, ABSENT FROM THE COMPANY’S OWN PRESS PAGE. The company has published two press releases, the April 2025 launch and the April 2026 round. A major outlet reports a Google deal that moved the former CEO and over a dozen staff. The company’s pages are silent. The finding is structural: a supplier of evaluation inputs changed its relationship to one of the labs it sells to, and the disclosure that exists is a newspaper’s reading of LinkedIn.

• Apr 17, 2025. Announced Mechanize, “backed by investments from Nat Friedman and Daniel Gross, Patrick Collison, Dwarkesh Patel, Sholto Douglas, and Marcus Abramovitch” (Mechanize).
• May 30 to Nov 25, 2025. Published eight essays, among them the automation-is-inevitable essay (Oct 6) and the reply to If Anyone Builds It, Everyone Dies (Nov 25) (Mechanize, essays).
• Apr 24, 2026. Announced $9.1 million at a $500 million post-money valuation (Mechanize).
• Jun 16, 2026. Posted the podcast episode “How bad data teaches models to write terrible code” (Mechanize, podcast).
• Aug 5, 2026. Business Insider reports Google in talks on a deal described as “$1.5 billion-plus” (Business Insider).
• Sep 11, 2026. Business Insider reports the deal completed, with terms undisclosed (Business Insider).

Alliance Map

Where ties on different axes run at the same time, the overlap is the finding. The company’s founders came from the nonprofit that builds benchmarks and measures the labs; the company sells environments and evals to labs; one of those labs is reported by a first-tier outlet to have completed a talent deal for it, with its former CEO and over a dozen former staff now at that lab. Each tie is stated by Mechanize, Epoch or Business Insider. They show an arrangement, not coordination.

Reach Assessment

Institutional: Mechanize’s environments are, in its own words, used by “frontier AI labs” to train and measure models, but no lab, no release document and no procurement standard names it. Memetic: its essays carry the argument that full automation is inevitable and desirable, and its reply to the AI-doom case is addressed to a named book; GBA Eval is its one public artifact.

Score basis. Atk 5 is reach: the working-here page says tasks “are bundled into environments, which are used by frontier AI labs either to train their models directly or to measure their capabilities” (Mechanize, working here). The points under a higher number are that no customer is named, no statute or release gate requires anyone to use it, and its public product is one benchmark (Mechanize, home). Peers: Epoch AI holds atk 5 on a published 987,000-user audience and cited use by central banks; Andon Labs holds 6 because three labs print its tests in their own release documents. Def 5 is funding insulation: the company says it “raised $9.1 million at a $500 million post-money valuation” from three named individuals and lists ten backers (Mechanize, press release; Mechanize, home). The points off are that revenue and customers are unpublished, that the buyers are the labs whose models the products train and score, and that the press page carries no account of the reported Google arrangement (Mechanize, press releases; Business Insider). Peers: Andon Labs holds def 5 with Y Combinator as its only named backer and undisclosed clients; Epoch AI holds 7 because it publishes donors of $70,000 or more and its paying clients. Hp 4 is entrenchment: two press releases in 18 months, no disclosed contract or installed base, and Business Insider’s report that over a dozen former employees are now at Google (Mechanize, press releases; Business Insider). The points under a higher number are the same absences. Peers: Andon Labs holds hp 6 on live deployments; Epoch AI holds 7 on installed data products.

Sources: Mechanize — home; Announcing Mechanize, Inc. — Mechanize; Mechanize raises $9.1M — Mechanize; Press releases — Mechanize; Press — Mechanize; Contact — Mechanize; Apply — Mechanize; Software Engineer — Mechanize; Research Engineer, Alignment — Mechanize; What working at Mechanize is like — Mechanize; How we interview software engineers — Mechanize; Essays — Mechanize; The future of AI is already written — Mechanize; Sweatshop data is over — Mechanize; Unfalsifiable stories of doom — Mechanize; Podcast — Mechanize; GBA Eval; Clarifying the creation and use of the FrontierMath benchmark — Epoch AI; Besiroglu on X — launch; Besiroglu on X — goal; Business Insider — Google in talks (Aug 5, 2026); Business Insider — Google completes deal (Sep 11, 2026).