The Mechanical Orchard story: how a company built on pair programming came to argue against code review
Mechanical Orchard sells one promise: it will rebuild the software running inside a large company, using AI to write the code, without breaking what is running. That promise is close to what we do, which is why we read the public record on it. This piece is built from 129 pages on the company's own site, six independent articles, two analyst pieces, two investor posts, and a marketplace listing. Every number that comes from the company or its investors is labelled as the company's own account, because a company's own site is a record of its claims.
One thing is worth saying at the top. The company was founded by the man who built Pivotal Labs, a firm whose reputation rested on pair programming and human judgment applied to every line. In July 2026 his own chief technology officer published an essay saying human review must stop. Watching a company gradually argue against its founder's method, in public, over four years, is the most useful thing in this record.
Part one: the founder and the idea
Rob Mee founded Pivotal Labs and was chief executive of Pivotal Software, which VMware bought for $2.7 billion in 2019. Both facts are reported by SiliconANGLE and TechCrunch in their own voice. Matthew Work, previously in leadership roles at Cognizant and Amazon, was named co-founder and chief operating officer in the company's February 2024 funding release. Kent Beck, who helped create Extreme Programming, joined as chief scientist.
The origin story is unusual, and Mee tells it himself in an essay on the company's site. In April 2020, at the request of a former director of the CDC, he assembled a team of former Pivotal people to build software for state governments handling contact investigation during the pandemic. They delivered it. Then they discovered the client could not run it: state IT departments were, in his words, "typically understaffed in IT and have many competing demands for their time and attention." So the team ran the software themselves.
That accident became the business model. Mee writes that "the model of building and running software turned out to be both efficient and fulfilling", and that the cycle of building, running and improving let them see feedback that a hand-over model hides. One company came out of that work first, Ratio PBC. Mechanical Orchard came second.
The founding idea, in his words, is a company "formed in the image of Pivotal Labs, but taking on the lessons of Pivotal Software and our recent response to the Covid pandemic". The problem he named was "legacy technical debt, uneven performance of software teams, and the burden of operating poorly written production systems". The phrase he chose for the work is worth keeping: "We perform digital transplants, duplicating the behavior of orphaned black-box systems, but in a form optimized for change."
Read that essay today and one thing is missing from it entirely: any mention of artificial intelligence. The company that now sells an autonomous code pipeline began as a company about operating software well.
Part two: the timeline
Dates from the company's own announcements index unless a source is named.
| Date | What happened | Source |
|---|---|---|
| November 2022 | Company launches, backed by Spider Capital, Bloomberg Beta and Cendana Capital | Mee's founding essay, company |
| 7 February 2024 | $24 million Series A led by Emergence Capital, reported at a $95 million valuation | Company release; SiliconANGLE citing Forbes |
| 28 February 2024 | First long independent profile, pilots described as six to eight weeks | Blocks & Files |
| 23 April 2024 | Edward Hieatt appointed to lead customer experience | Company release |
| 6 August 2024 | $50 million Series B led by GV | Company release; TechCrunch |
| 15 November 2024 | Kent Beck announces he is stepping down as chief scientist | Beck's own newsletter |
| 3 April 2025 | Imogen launches as a named platform, with Thoughtworks as first partner | Network World; company |
| 3 June 2025 | Named a Gartner Cool Vendor in AI Code Assistants | Company release |
| 24 November 2025 | Listed on Google Cloud Marketplace | Company release |
| 8 December 2025 | Listed on AWS Marketplace | Company release |
| 9 April 2026 | Available to public sector buyers through Carahsoft's GSA schedule | Carahsoft |
| 22 April 2026 | SulAmerica presents at Google Cloud Next 2026 | Google Cloud Next session listing |
| 7 July 2026 | Chief technology officer publishes "The Software Dark Factory" | Company |
| 9 July 2026 | Summer release: autonomous build and verification pipeline, Leidos partnership | Company release |
| 13 July 2026 | Analyst pushes back on the verification claims | HyperFRAME Research |
The date we find most telling is 15 November 2024. Kent Beck, who is the reason the Extreme Programming lineage was in the company's story at all, stepped away as chief scientist. He wrote kindly about it in his own newsletter, writing "I believe in the mission of MO. I believe in the people there", and giving a personal reason: he wanted his remaining time for his own priorities. Twenty months later the company published an essay saying code must not be reviewed by humans. We are not suggesting those two facts are connected, and nobody has reported that they are. They simply happened before and after the company's change, and Beck's name is the one most associated with the practice being set aside.
Part three: the money
Two rounds have been announced. Everything else about the money is contested, including by the company itself.
| Figure | Value | Date | Who said it |
|---|---|---|---|
| Series A | $24 million, led by Emergence Capital | Feb 2024 | Company, confirmed by SiliconANGLE and the law firm Fenwick |
| Series A valuation | $95 million | Feb 2024 | SiliconANGLE, citing Forbes |
| Revenue, end of 2023 | "more than $10 million annually", profitable | Feb 2024 | SiliconANGLE citing Forbes; Emergence Capital |
| Series B | $50 million, led by GV | Aug 2024 | Company, confirmed by TechCrunch |
| Series B valuation | Never disclosed | ||
| Total raised | $74 million | Aug 2024, repeated Apr 2025 | TechCrunch, then Network World |
| Total raised | "over $84m" | Apr 2025 | The company, in the Thoughtworks joint release |
| Total raised | "Over $80 million" | Jan 2026 | Rob Mee, in a Pulse 2.0 interview |
Two things follow from that table. The first is arithmetic: $24 million plus $50 million is $74 million, which is exactly what both independent outlets reported. The company's own larger figures imply money raised that was never announced, and its two larger figures do not agree with each other either. Aggregator sites carry totals of $93.2 million and $102 million; those pages refused to load for us, so we are not publishing them as facts.
The second is a gap. No revenue figure has been published by anyone since February 2024. A company that told Forbes it passed $10 million and was profitable in its first full year has said nothing about revenue in the nineteen months since, through two marketplace listings, three partnerships and a major product release.
The headcount figures show a pattern too, and it is worth naming.
| Reported headcount | Date | Who said it |
|---|---|---|
| "50+" | 7 February 2024 | Company |
| "More than 50" | 7 February 2024 | SiliconANGLE |
| "80+" | 6 August 2024 | Company |
| "approximately 90" | 7 August 2024 | TechCrunch |
| "about 100" | 3 April 2025 | Network World |
| "100+" | 6 January 2026 | Rob Mee |
The company says "80+" and a reporter with access to the same company says "approximately 90" the next day. That is the ordinary gap between a marketing minimum and a reported count, and it is the reason a figure with a plus sign after it should be read as a minimum. Note also the pattern: roughly 50 to roughly 100 over two years, for a company that raised $50 million and sells a platform. It is not hiring like a firm that bills for people.
Two finance terms, in plain words, because they appear above. A valuation is what investors agreed the whole company was worth at the moment they put money in, and it is a negotiated number. Profitable means the money coming in covered the money going out over that period, which for a company two years old is unusual and is a real signal.
Part four: the product
For the first two years there was no named product. Blocks & Files described the work in February 2024 as an engagement with close personal involvement: pilots of six to eight weeks, then component-by-component migration, with old and new run side by side until the outputs matched. Mee told that outlet there is "no silver bullet, and never will be" for moving legacy applications to the cloud in bulk.
The product got a name on 3 April 2025: Imogen. Its platform page describes three modules, and these are the company's own words. Analyze "reads your source directly to map how the system works and what it depends on". Rewrite "generates modern code and verifies it against real production data, slice by slice". Orchestrate runs legacy and modern components side by side, so buyers can "modernize in order of business, one slice at a time, with no big-bang cutover".
The technical claim underneath all three is behavioural equivalence, and the company's field chief technology officer, Sam Sanders, sets it out in a research paper on the site. He defines success as "behavioral equivalence", meaning the new system produces identical output to the old one for the same inputs, and argues that a language model translating COBOL to Java will produce code that compiles while missing file organisation, numeric encoding and error codes. His proposed answer is to feed identical inputs to the old system and the new one and compare, treating the running legacy system as the correct reference.
That is a strong idea, and it is worth spelling out why. In most old software systems the original specification is gone, the people who wrote it have retired, and in some cases the source code itself has been lost. What survives is a system that runs. Treating that running system as the specification is the only honest option available, and it gives the rewrite something to be checked against that cannot be wrong about its own behaviour.
The homepage puts the same point in one line: "Both versions compile and run; only one is verifiable."
Then came July 2026, and the product's biggest change. The summer release added what the company calls an autonomous build and verification pipeline, which it also calls a software dark factory, removing manual code review from the process. The release says customers are "regularly exceeding 10,000 lines of verified code per engineer per week, with the potential to gain further speed by an order of magnitude". That is the company's own figure and no outside party has tested it.
The sentence in that release that explains the whole company is this one: code moves through the pipeline "as fast as it can be verified". The limit on delivery speed, in their design, is the checking system.
Part five: how it sells
There is no price on the company's website. There is no product screenshot on the homepage, the platform page or the government page. What there is instead is a research report, a cost calculator, a 129-page content library, and a free proof of concept.
That last one is the commercial mechanism, and it is well built. The proof of concept page offers to run Imogen against a real slice of a buyer's codebase "at no cost, under NDA, without access to your data", delivering a modernization roadmap, an initial Java replica of about 100,000 lines with test harnesses, and a demo environment. The page opens by saying that "Most modernization decisions get made on faith". For a buyer who has been harmed by a failed migration, being handed evidence about their own system before signing anything is a strong offer.
Selling happens through partners more than through a direct salesforce. Thoughtworks was announced as the first partner on 3 April 2025, pairing its system-comprehension tooling with Imogen's rewrite. Leidos followed in July 2026 for US federal work. Carahsoft added Imogen to its GSA schedule in April 2026, which lets government and education buyers purchase through an existing contract. The platform is listed on both Google Cloud Marketplace and AWS Marketplace, and integrates with AWS Transform and Google's migration tooling.
One public price does exist, and we verified it ourselves in the listing data. The AWS Marketplace listing for Imogen carries a one-month contract for a single unit at $100,000. The description of that unit reads "Reach out to us for pricing details." A buyer can therefore see a six-figure monthly number and, in the same listing, be told to ask for a price.
The channel this company lacks is a self-serve one, and that is deliberate. Nothing here can be bought without a conversation.
Part six: the customers
This is the weakest part of the record, and the weakness is itself the finding.
| Customer | Where | Relationship reported by | Result, by the company's own account |
|---|---|---|---|
| Omni Logistics | US | Named in Feb 2024 funding materials; repeated by SiliconANGLE and Blocks & Files, both writing from the same announcement | Core applications moved to cloud |
| SulAmerica | Brazil | Company media page and the Google Cloud Next 2026 session listing | Over 130,000 lines of COBOL taken to verified equivalence |
| "Fortune 20 Automotive Company" | Not stated | Company homepage | An 18-month effort delivered "in approximately five months" |
| A global manufacturer of industrial equipment | Not stated | Company case study | Four billing jobs migrated and running in production within four months |
| "Fortune 500 Retailer" | Not stated | Company case study | Prior attempt cost $40 million over three years; first cutover 27 June 2023 |
| "Fortune 100 Healthcare Company" | Not stated | Company homepage testimonial | No result given, praise for rigour |
| US Navy | US | A video title on the government page | No contract or agency confirmation found |
Two observations about the shape of that evidence.
The first: in almost five years there is no named customer engagement confirmed by anyone with no interest in it. Omni Logistics comes closest, and both outlets that mention it were writing up the same funding announcement, so that is one source repeated three times. SulAmerica is the strongest case, because a named person from the customer stood on a stage and said it, but the stage was Google Cloud Next, and Google is both a partner and, through GV, the lead investor in the company's largest round. That is not neutral confirmation.
The second: the company names its customers by Fortune tier, which is a choice that gives up some credibility so that more customers can be mentioned. "Fortune 20 Automotive Company" is a set of about four firms, so the label is close to a name without being one. Large enterprises routinely forbid suppliers from naming them, so this is a normal constraint. It does mean that a buyer's only real check is the free proof of concept, which is perhaps why the company invested in making that offer good.
The US Navy line deserves a direct statement: a video titled "Modernizing mission-critical systems without breaking the navy" sits on the government page, and we found no contract, no agency confirmation and no independent report. Nobody should read that page as the Navy being a customer.
Part seven: what investors say
Emergence Capital led the Series A and published its own reasoning on the same day. These are an investor's words about its own portfolio company, so treat them as a description of the investment. Emergence supplied the market framing the company still uses: $13 trillion of the US economy dependent on mainframes, 71 percent of the Fortune 500 still running them, 44 of the top 50 banks, 10 of the 10 top insurers. It also supplied the revenue detail, "double digit millions in recognized revenue" and profitability at the end of 2023.
GV led the Series B. Crystal Huang, a general partner there, said the team "impressed us deeply with their AI-powered approach to enterprise application and infrastructure modernization", and named what GV was buying: understanding of complex enterprise IT, experience scaling teams, a roadmap, and market size. Mee told TechCrunch that "The raise was unsolicited", which is a claim by the person raising the money and cannot be checked.
The example both investors are relying on is Pivotal. Kent Beck, writing on the company's site, recounts that "more than 3000 engineers, designers, and product folks became Pivots" and that Pivotal reached a market value above a billion dollars. The investment case is that the person who built one large software services business from a method can do it again with a better method.
What nobody has published is a valuation for the Series B or anything after it, and the investor list has changed. The February 2024 release named six firms. The 2026 boilerplate names four, of which two, Fin Capital and MongoDB, appear in no announced round on the company's own site.
Part eight: what the press says
The independent coverage is limited, concentrated in two funding weeks, and one of the important pieces cannot be read.
Forbes first reported the Series A on 7 February 2024 and is the origin of the $95 million valuation, the $10 million revenue figure and the profitability claim. Its page returned an error for us and for all three of our researchers. Everything anyone knows from it, including this piece, arrives through SiliconANGLE citing it, and this piece says so at every use.
TechCrunch covered the Series B and is the source for "approximately 90 employees" and for the $74 million total. It also carries the single most important sentence in the whole record for our purposes, and it is the reporter's paraphrase: that the company always has developers involved to debug and review the finished product. We come back to it in part ten.
Blocks & Files wrote the deepest independent profile in February 2024, and it is the only piece that describes the delivery model concretely: six to eight week pilots, component-by-component work, old and new run side by side. It also noted the model looked like a traditional consulting engagement.
Network World covered the Imogen launch in April 2025 and is the source for "about 100 employees" at that date.
Every press item we found is in English. There is none in Portuguese, despite the company's most prominent named customer being Brazilian, and none in Italian or German, despite reported offices in both countries. For a company selling to global enterprises, the absence of any local-language coverage is a real limit on how widely its story has spread.
Part nine: what critics and analysts say
The honest summary: almost nobody has attacked this company in public. We searched for lawsuits, regulator actions, layoffs, failed projects and customer complaints, and found none. That describes the state of the published coverage, and a company can go unexamined for years without that meaning anything either way.
The one substantive skeptic is Stephanie Walter of HyperFRAME Research, and her two pieces are precise. In May 2026 she wrote that "Translating code syntax without mapping operational behavior merely shifts systemic debt from an on-premises box to an expensive cloud environment", and warned that running a modernized cloud component alongside a legacy core "introduces material operational friction", which is directed at the Orchestrate module.
In July 2026 she challenged the verification claim itself: "Behavioral proof is only as strong as the behavioral coverage." Her full objection is the sharpest thing written about the company by anyone: "A synthetic replica of production flows can reduce migration risk, but it cannot magically guarantee that every edge case, seasonal transaction pattern, peak-load anomaly, or undocumented exception path has been captured."
One caveat, and it matters. Neither page carries a disclosure of whether the company sponsors HyperFRAME's coverage, and the framing is friendly to the company's positioning. The criticism is real; the independence is unconfirmed.
Gartner named the company a Cool Vendor in AI Code Assistants in a report published 3 June 2025. That report is behind a paywall and we did not read it. A Cool Vendor listing is a recognition, and Gartner's own disclaimer, which the company reproduces, says Gartner "does not endorse any vendor, product or service depicted in its research publications". It should never be described as a ranking.
On anonymous reviews: Glassdoor shows roughly two to three reviews, PeerSpot shows zero, and there is no presence on G2. Three reviews of a company of about a hundred people measures nothing, and we are giving the count and stopping.
Part ten: what the people inside say
This is where the record gets interesting, because the company's engineers are arguing with their own past position in public.
On 7 July 2026, chief technology officer Roberto Ostinelli published "The Software Dark Factory". Its central claim is two sentences long: "Code must not be written by humans. Code must not be reviewed by humans." His argument places the care in the automated process: "Build a process so well-specified and so well-checked that human hands in the middle would only slow it down or introduce error." His condition is explicit: "Where there's a definition of done and a way to prove you've reached it, you can take humans out of the loop without taking judgment out of the work." And his warning about doing it badly is the best line in the essay: "Autonomy without boundaries is reckless; autonomy within boundaries is the entire proposition."
Compare that with what TechCrunch reported in August 2024: that the company always has developers involved to review the finished product. Two years, two speakers, and a reversal that the site never acknowledges anywhere.
Fifteen days after Ostinelli, Rachit Awasthi published "Illuminating the dark factory", which makes the position less strict. His version keeps people watching the process while removing people from doing the work, and his criticism of code review is worth quoting because it is a criticism of a practice this company was founded on: review is "a human skimming a diff, approving on reputation and fatigue as much as rigor". The principle he reaches is the one we agree with most: "The worker is never its own inspector: every piece of work meets an independent check, and work that fails is sent back until it passes."
Set against that, the engineering job posting still describes pair programming as "Mechanical Orchard's default mode of working together", and lists test-driven development, refactoring and Extreme Programming practices. The company's stated values are three sentences: "Do the right thing. Do what works. Be kind." No salary range is published, and the board listed three open roles when we read it.
So inside one company, in one quarter, you can find an essay saying humans must not review code and a job advertisement recruiting people to pair-program. Both are true. The pairing is how the pipeline gets built; the pipeline is what removes the review.
Part eleven: the details that do not match
Everything here we verified in the page source ourselves.
- The platform page's scrolling list of supported technologies reads "OL" where COBOL belongs, in both copies of the list. The word COBOL appears correctly 29 times elsewhere on the site. The company's main product page misspells the language it is best known for rebuilding.
- The same list reads "IDMS ADS" in one copy and "ADMS ADS" in the other.
- Webflow placeholder text is live on 12 of the 23 pages we downloaded, including funding announcements: a comment block containing "Lorem ipsum dolor sit amet", an "Author Name" and a "Comment Time".
- Mee's founding essay carries a page date of 3 April 2022 but describes a November 2022 launch and says "Launching today". The page is dated seven months before the event it announces.
- Edward Hieatt is "Chief Customer Officer" in the April 2024 release and in the research report, and "Chief Operations Officer" on the about page. The older pages were never updated.
- Matthew Work is named co-founder and chief operating officer in February 2024 and does not appear on the about page in September 2026. We found no announcement, so we are calling this an absence.
- The contact page shows two different San Francisco addresses at once, on Jackson Street and on Post Street, with nothing distinguishing an office from a registered address.
- The warranty project is "four months" on the case study and in an April 2026 post, and the same page calls it an "80% acceleration". Eighteen months to four is 78 percent. The homepage separately describes a different customer at "approximately five months".
- The company's own cost calculator credits its comparison baseline to "Vikas Pujar's LinkedIn post" and a "Rocket Software survey", with no link and no date, while the figure for its own process, 1,200 lines per developer per day against a traditional 150, is uncited.
- The Series B release says "as many as 75%" of modernization projects fail to complete. Emergence Capital's post, the source of most of the company's other market figures, says 70 percent.
We list these because our own site is checked against a specification before it is released, and a marketing page is a change like any other. A company whose entire argument is that unverified output should not reach production has put placeholder text into production on twelve pages.
Part twelve: the open questions
- What is the behavioural coverage? The verification argument depends on what share of real system behaviour the captured data flows represent. The metric that would answer it: the percentage of production transaction paths exercised during verification, with the measurement window stated.
- What happens at the first disagreement? When the old system and the new one produce different output, something has to decide which is correct, because legacy systems contain bugs that downstream systems depend on. The metric: the count of behavioural differences found per workload, and the share resolved as "keep the old behaviour".
- Did the review actually stop? The July 2026 release says the pipeline removes manual code review, and a colleague's essay fifteen days later keeps human eyes on the process. The metric: the share of generated code that reaches production with no human having opened it.
- What does an engagement cost? One marketplace listing shows $100,000 per month for one unit and tells the buyer to ask. The metric: total contract value for a completed workload, against the mainframe spend it removed.
- Where did the money beyond $74 million come from? Two announced rounds add to $74 million and the company says over $84 million. The metric: a named round, date and amount, or a correction.
What we take from it
The useful idea here is not AI-written code, which everyone now sells. It is that the check has to be independent of the thing being checked, and that when a running system exists, it is a better specification than any document about it. Reveneau's own method has the same structure: the code is written by AI, and it has to pass an eval suite written from the specification before it is released. We have not measured our own results against anyone's, and nothing on this site claims we have.
The second thing we take is a warning about our own argument. This company published an essay saying human review must stop, and a colleague published a gentler version two weeks later, and neither the essay nor the site acknowledges the position it replaced. The honest way to change a position in public is to say you changed it. We have changed ours, from accountability resting on a named reviewer to accountability resting on the eval suite and on the company, and this is us saying so.
The third is the question to ask in any vendor conversation, ours included: what is the work checked against, and what separates the checker from the writer? A supplier who cannot answer that in one sentence is selling speed with no control on it.
Credits
This piece is based on reporting by Kyle Wiggers at TechCrunch, Chris Mellor at Blocks & Files, Michael Cooney at Network World, Mike Wheatley at SiliconANGLE, Amit Chowdhry at Pulse 2.0, and Alex Konrad at Forbes, whose Series A story is the origin of figures everyone else repeats. The sharpest analysis comes from Stephanie Walter at HyperFRAME Research. Credit is also due to Mechanical Orchard itself, which publishes enough of its own thinking, including thinking that contradicts its earlier thinking, to be read closely. A company that argues with itself in public is easier to trust than one that never changes its mind.
Every rewrite needs a standard to check it against.
Sources
Independent reporting
- TechCrunch, 7 August 2024: the Series B, the $74 million total, "approximately 90 employees", and the 2024 position that developers are always in the loop.
- Blocks & Files, 28 February 2024: the delivery model, pilot length, offices, and "no silver bullet, and never will be".
- SiliconANGLE, 7 February 2024: the Series A, the investor list, and the Forbes-sourced valuation and revenue figures.
- Network World, 3 April 2025: the Imogen launch and "about 100 employees".
- Kent Beck's newsletter, 15 November 2024: his own account of stepping down as chief scientist.
- Forbes, 7 February 2024, by Alex Konrad: the origin of the $95 million valuation and the $10 million revenue figure. The page would not load for us, so every use of those figures here is attributed to SiliconANGLE citing Forbes.
Analysis
- HyperFRAME Research, 22 May 2026: the operational-friction objection to running old and new side by side.
- HyperFRAME Research, 13 July 2026: "Behavioral proof is only as strong as the behavioral coverage."
- Gartner, Cool Vendors in AI Code Assistants, 3 June 2025. Paywalled and not read; cited only as the company reports it, with Gartner's own non-endorsement disclaimer.
Company and investor statements, all self-declared
- The Seeds of Mechanical Orchard: Rob Mee's founding essay, the pandemic origin and the founding thesis.
- The platform page: the three modules, the supported technologies, and the two typographical errors noted above.
- Mainframe modernization is a verification problem: Sam Sanders on behavioural equivalence.
- What's new in Imogen, the Summer 2026 release: the autonomous pipeline, the Leidos partnership, and the 10,000 lines figure.
- The Software Dark Factory: Roberto Ostinelli on removing human review.
- Illuminating the dark factory: Rachit Awasthi on independent inspection.
- The Series B release, 6 August 2024: the GV round, "80+", and the 75 percent failure figure.
- The Gartner Cool Vendor release, 9 June 2025: the recognition, the $66 billion market figure, and Gartner's disclaimer.
- The proof of concept offer: what a buyer gets before signing.
- The manufacturing warranty case study: the four-month claim in a customer's words.
- The AWS Marketplace listing for Imogen: the only public price, verified in the listing data on 2026-09-17.
- Emergence Capital, 7 February 2024: an investor writing about its own portfolio company, and the origin of the market figures the company still uses.
