Reason Machines (YC Spring 2026): teardown vs XMarts

In one line: Reason Machines sells Reason Agent, a coding agent that takes a task, tests its own work and opens a pull request with proof, from the web, the terminal, a Mac app or Slack [yc] [site] [press]. Threat: 2 because it is the same buyer and problem as W1, but it is two people, four months into this product, with no named customer [yc] [site]. Expect a YC partner to name it anyway: it is in the same batch pool as us, it is live, and it competes head on with Codex, which is one of the engines we run on [blog].

1. Fact file

Data conflicts to know: Product Hunt says "Launched in 2025" while YC says founded 2026 [press] [yc]. The YC list shows a former name "EasyClaw"; I found no source tying EasyClaw to this company, so that name is UNVERIFIED and left out of the Path.

2. Their proposition, reverse engineered

Check on their numbers: a search of the independent Artificial Analysis index found Codex listed at 67 on a different model setting, not 62.5, and no entry for Reason. Their 65.1 is self-reported and unverified [press] [blog]. Their benchmark name "Terminal Bench 4.0" is also not the name that index uses [press].

3. Positioning

• Category they claim: "the world's best coding agent"; "Autonomous Software Engineer" [site].
• Category a YC partner would put them in: coding agent harness, competing with Codex and Claude Code on cost and score (inference from [blog]).
• What they want to be known for: the cheapest way to get a good result from a coding model.
• Where they sit on the map: layers L1/L2/L3 = 1/2/2, buyer = size unknown, developer.

4. Three layers vs ours

5. Same and different

6. Wedge overlap

Not a fit, but watch: W8 (skill upkeep from accepted and rejected work). Reason does not claim it today [blog]. The nearest rival for the W1 plus W8 angle stays Nessie (brief). W2 (is my agent fleet producing?): Reason's site shows a sample spend dashboard in a demo, which is not a product claim [site].

7. Whitespace (where they stop)

• Layer gap: the product stops at a checked pull request. Nothing public says an accepted or rejected PR changes how the next run works [blog] [site]. To own it we need a rule-change log tied to checked runs, with rework falling, on one customer's code.
• Buyer gap: not clear. They say "engineering teams" but show no team features, no logos and no sales role [site] [yc] [jobs]. To own it we need one named small startup team with the pain on file. Do not call this a gap yet.
• Number gap: they count benchmark score and cost per solved task [blog]. They show no merged-PR rate, no rework rate and no hours saved for a real team. To own it we need one team's number: share of agent PRs merged without rework, before and after. Note our own number is weak today (about 15% of jobs reached main).
• Price gap: price unknown [press] [site]. Free local use sets a very low floor. We can not win by being cheaper per task; they lead on that claim (they claim).
• Speed gap: not a gap. They have a free sign-in and a local mode [press] [site]. Our system is Mac-bound and onboarding is unproven.

8. For the YC application

• Proven today: Marton's logs show his system has run about 5,600 jobs across agents from three makers, each checked by a different AI and then automatic tests, with about 160 failure records that each added an automatic check. Reason's public proof is a benchmark and a cost per task, and the site names no customer [blog] [site].
• Draft answer to "how are you different from Reason?": DRAFT. "Reason makes one agent cheaper per task. Our system runs agents from three makers, has a different AI check each job, and writes a record and a check for every failure." Founders say it their own way.
• Future claim (needs proof): "Our agents' work gets merged more often, because every rejection becomes a rule." Needs one startup team's repo run for 4+ weeks, with the share of agent PRs merged without rework before and after.
• What we understand that they don't: the hard part of 24/7 coding is trusting the output and keeping the line moving, not the cost of a task. proven only as a lived experience: Marton's logs show cost blowouts, silent stops and a 15% merge rate. needs proof that a team feels the same pain.
• Copyability (M3): medium. They already open PRs with proof, orchestrate sessions and work across several models [site] [press], so a cross-check by a second model is a feature for them (inference). Learning from accepted and rejected PRs needs weeks of real customer data, which they have not shown [blog]. Chasing it fits a harness company less than cutting cost does (inference). Not hard, so M3 can not score above 3 against Reason alone.
• Why now (M4): None found in this teardown. Their own timeline (a new model generation, a rename in October) is general AI progress, not a dated change that helps us more than them.
• Do not claim: "more trusted than Reason" (no side by side test). "Our agents ship reliably" (about 15% of jobs reached main). "Cheaper than Reason" (we have no per-task cost number). Anything that calls the market empty: Reason is live, in YC and ranked #8 on its Product Hunt day [press].

M2 (what YC funds now) is NOT judged here. It comes from the themes of many companies in Pass 3.

9. Ideas worth taking

1. Make the proof part of the output: a short video and before and after screenshots on every PR [press]. A buyer sees the work without reading code.
2. Publish the benchmark run with public job links, so anyone can check the number [blog]. Use the same shape for our own number once we have one.
3. One plan hits a limit, work moves to another plan [press]. Marton's logs show we already do this; it is a good one-line demo.

10. Unknowns

• Who pays, and how much. Look for a pricing page after sign-in, and ask a user.
• Real users. One search summary mentioned a traction claim from a founder's social post; its source was a LinkedIn page, which I did not open. Treat as UNVERIFIED.
• Funding: $125K vs $500K. Check the YC deal terms for Spring 2026 or a founder post on X.
• Whether the Ara-era "stored context" still exists in Reason. Try the product, or read the changelog.
• Whether the Reason 1.1 numbers hold. Look for an Artificial Analysis entry or the public GitHub runner named in the post.
• SOC 2 status. The trust page returned only its title, so Type 1 or Type 2 is unknown.
• Whether they sell in the UK. Nothing found.

Sources

• reasonmachines.com [site]
• Reason Agent 1.1 research post, 10 Oct 2026 [blog]
• YC company page [yc]
• YC page under the Ara name [yc]
• Product Hunt launch page [press]
• RuntimeWire on Ara's July update, 9 Aug 2026 [press]
• Standout job board listing [press]
• Dealroom profile [press] (page blocked a direct read; used only via a search summary)
• Artificial Analysis coding agent comparisons [press]
• YC jobs listing on the company page [jobs]
• Trust center (title only readable)