Replicas (YC Spring 2026): teardown vs XMarts

In one line: Replicas runs Claude Code, Codex and other coding agents in cloud machines. Engineers hand them tickets from Slack, Linear or GitHub and get pull requests back [site] [yc]. Threat: 3 because it is the same buyer and the same problem as W1 (24/7 coding for startups), it raised a $4M seed on 13 Oct 2026, and it has named customers [blog] [site].

1. Fact file

2. Their proposition, reverse engineered

3. Positioning

• Category they claim: Cloud coding agent; background coding agents; harness-agnostic cloud infrastructure for agents [site] [blog].
• Category a YC partner would put them in: developer tools, background coding agents, in a crowded category (inference).
• What they want to be known for: the best computer for any coding agent, wired into the tools a team already uses.
• Where they sit on the map: layers L1/L2/L3 = 2/2/2, buyer = sme (named customers are startups), developer.

4. Three layers vs ours

5. Same and different

6. Wedge overlap

7. Whitespace (where they stop)

• Layer gap: they stop at a PR a human must still approve and merge [blog]. Owning the merge decision means showing a gated path that merges without a person on the routine cases. We would need a run of real merges with a low bad-merge rate on a customer repo.
• Number gap: their proof is usage and share, and their one honest test shows 12% of PRs wrong or abandoned [blog]. We would need a before/after number a head of engineering would sign: merged PRs per week and rework, over 4+ weeks.
• Buyer gap (weak): they already sell to startups of all sizes and are opening an enterprise tier [site]. Utility Profit's case study title says "all departments" (page not readable, UNVERIFIED), so non-engineers may come next. No true buyer gap found for W1.
• Price gap (none): from $50 a month with $100 free [site]. We cannot undercut them.
• Speed gap (none): self-serve sign-up and free credit; first value in minutes [site].

8. For the YC application

• Proven today: Marton's logs show his system runs about 5,600 agent jobs across three makers' engines. Each job is built in its own copy of the code and checked by a different AI before automatic tests. About 160 failures each got a record and an automatic check. A May 2026 Replicas post says nothing in its stack then knew which PRs to auto-merge and which to escalate [blog]; they may have changed this since.
• Draft answer to "how are you different from Replicas?": DRAFT. "Replicas gives agents a computer and hands you pull requests to review. We care about what happens after: checking the work with a different AI and deciding what is safe to merge. We have run that loop on our own business for five months." Founders say it their own way.
• Future claim (needs proof): "Teams merge agent work without reading every PR, and we show the bad-merge rate." Needs one customer repo run for 4+ weeks with a before/after number (merged PRs, rework, bad merges).
• What we understand that they don't: the hard part of 24/7 agent coding is trusting the output, not producing it. needs proof on a customer; Marton's logs show the failure history on his own system, including that only about 15% of jobs reached main.
• Copyability (M3): easy. They already run several engines, switch an agent from Claude to Codex mid-task [blog], and used a separate review tool in their own test [blog]. A second-engine review step is a feature for them, and it fits their buyer (inference). Memory, analytics and a "manager" direction show they are moving toward the same loop [blog]. The deeper part (a failure taxonomy and a gated merge path built from real incidents) would take them longer, but nothing public shows they want it yet. Not hard, so M3 cannot score above 3 against Replicas.
• Why now (M4): None found in this teardown. Do not use "cloud agents became possible" or the Stripe and Ramp background-agent posts: they help Replicas as much as us.
• Do not claim: "trusted" or "safe to merge" as a result we have already delivered (only about 15% of jobs reached main, and landing was fragile in October). "We have customers" or "we are cheaper". "Replicas cannot check work": they used a review tool, and they may add one. Any claim about their revenue.

9. Ideas worth taking

1. Publish an honest outcome table on your own backlog (50% mergeable first time, 12% wrong or abandoned). It builds trust, and we can copy the shape using our own jobs.
2. Make setup a saved, shared environment per customer, so the second task starts warm. It cuts time to first value.
3. Start the trigger where the work already lives (Slack, Linear, a PR comment), not in a new app.

10. Unknowns

• Real revenue, number of paying customers and how much of the 7M minutes is paid or free credit. Ask a customer or read their next update.
• Whether they plan a merge gate or an outcome score. Watch their blog and changelog for "auto-merge", "review" or "quality".
• Where their buyers are based. No UK signal found. Check customer pages and careers.
• Utility Profit's case study and several blog pages did not load in full; the memory and analytics posts were read in summary only. Re-read the full text before quoting.
• A "V3" label appeared in a search result but not on their site (UNVERIFIED).
• Team size: YC page says 2; a June 2026 review said 1. Newer hires are possible.

Sources

• replicas.dev [site]
• Pricing [site]
• Customers [site]
• Enterprise [site]
• YC company page [yc]
• YC jobs page [jobs]
• Announcing our seed round, 13 Oct 2026 [blog]
• Replicas V1, 18 Mar 2026 [blog]
• What Replicas is doing in 2026, 5 Jan 2026 [blog]
• We gave Replicas our entire backlog, 16 May 2026 [blog]
• Giving agents memory, Sep 2026 [blog]
• Harness-agnostic analytics, 23 Aug 2026 [blog]
• Why cloud coding agents should be harness agnostic, 10 Sep 2026 [blog]
• Fallbacks and forks, 4 Sep 2026 [blog]
• Refreshing the Replicas brand, 12 Aug 2026 [blog]
• Blog index [blog]
• E2B customer story: Replicas, 27 Aug 2026 [press]
• Replicas review, rywalker.com, June 2026, third party [press]
• Seed round listing, TradedVC, read only through a search summary (page returned 403); used only to cross-check the blog [press]