Lemma (YC Fall 2025): teardown vs XMarts

In one line: Lemma reads the traces of AI agents that a startup runs in production, finds the failures that throw no error, groups them into issues, alerts Slack, and hands the fix to the team's coding agent [site] [docs]. Threat: 3, because it uses our W2 words ("silent failures"), sells to the same engineering buyer, is funded, live and names customers [site] [press]. The one soft spot: it watches agents that startups ship to their own users, not the agents that run a company's own work (inference, see section 7).

1. Fact file

2. Their proposition, reverse engineered

3. Positioning

• Category they claim: production monitoring for AI agents; "adaptation layer"; continuous learning [site] [blog] [yc].
• Category a YC partner would put them in: agent observability, in the same pile as Langfuse and Arize but pointed at meaning, not errors (inference).
• What they want to be known for: the tool that finds the agent failures nobody knew to look for, and fixes them.
• Where they sit on the map: L1/L2/L3 = 1/1/1. Buyer: sme, developer.

4. Three layers vs ours

5. Same and different

6. Wedge overlap

7. Whitespace (where they stop)

• Layer gap: they check and propose; they do not run the work [site] [docs]. To own it we need to show agents doing a customer's work, checked, with the monitor built in.
• Buyer gap: no named customer is an agent fleet doing a firm's internal work (Janet AI and Zarna may be close; UNVERIFIED) [site]. To own it we need one team running coding or ops agents who say silent stalls cost them days.
• Number gap: no public result in money or hours saved; only volume and a 90% iteration claim [site] [yc]. To own it we need a before/after count of silent failures caught and days saved on a real fleet.
• Price gap: none found. They have a trial and checkout [docs].
• Speed gap: none found. They claim setup under 10 minutes [site].

8. For the YC application

• Proven today: Marton's logs show his system ran about 5,600 agent jobs over five months and kept a heartbeat that reports what is stuck, what silently stopped and what it costs. About 160 failures each got a record and an automatic check. Lemma's public proof is volume of traces and logos, not work done [press] [site].
• Draft answer to "how are you different from Lemma?": DRAFT. "Lemma watches the agents a startup ships to its users. We watch agents that do the company's own work, and our own fleet has run that way for five months." Founders say it their own way.
• Future claim (needs proof): "We catch the stalls and silent failures in a team's agent fleet that trace tools miss." Needs one outside team's fleet watched for 4+ weeks, with a count of failures caught and a before/after number, plus the heartbeat decoupled from the relay-queue schema (inventory).
• What we understand that they don't: in a fleet that does work, most silent failure is in handover, supply and landing, not in a bad answer. needs proof outside Marton's own fleet; Marton's logs show silent-success and landing faults recurring many times.
• Copyability (M3): easy. Their traces already accept any agent through OpenAI Agents, Anthropic and LangGraph, and they just added checks on "completed tasks" (Oct 2026) [site] [docs]. Watching a coding fleet is another integration. They also hire a forward deployed engineer to chase customers [jobs]. Not hard, so M3 can not score above 3 against Lemma.
• Why now (M4): None found in this teardown. Do not use "agents are going to production": that is general AI progress. The Gartner "40% of agent projects cancelled by 2027" line sits in a contributor article, so it is not a clean source [press].
• Do not claim: "nobody watches for silent failures" (they do, in public). "Lemma only watches chatbots" (their customers include always-on agent makers [site]). Any customer result for W2.

9. Ideas worth taking

1. Install by pasting a setup prompt into the buyer's own coding agent [site]. It makes onboarding feel like minutes.
2. Every alert carries a confirmation status and proof, so people trust it [docs]. Our heartbeat could show the same.
3. After a fix ships, auto-create a check that watches for the same failure [site]. That matches our "failure gets an automatic check" habit, shown to a buyer.

10. Unknowns

• Price and plan limits. Look in the dashboard (platform.uselemma.ai) or ask a user who has a quote.
• Whether they price by trace volume (inference only).
• Whether the $2.3M and 1M traces a day are exact. Both come from their own press release [press].
• Whether any customer runs an internal ops or coding fleet. Look at the case studies behind /demo and the Janet AI and Zarna pages.
• Whether they track cost per agent. Check docs for spend or token metrics.
• UK or European customers. Nothing found; the roles ask for US work authorisation [jobs].
• Real team size: the blog lists at least one more author than the 2 YC counts [blog].

Sources

• uselemma.ai [site]
• Lemma docs home [docs]
• Lemma changelog, entries to 15 Oct 2026 [docs]
• Lemma blog index [blog]
• Introducing Lemma, 24 Apr 2026 [blog]
• Lemma careers [site]
• YC company page [yc]
• YC job ad: Product Engineer [jobs]
• Lemma $2.3M pre-seed, press release via VCA Online, 13 Aug 2026 [press]
• VentureBeat contributor article on Lemma, 8 Oct 2026, contributor content [press]