Nessie (YC Fall 2025): teardown vs XMarts

In one line: Nessie saves every AI chat, agent run and meeting into one searchable library, lets a team share it, and now tracks and edits the agent skills built on top of it [site] [yc]. Threat: 2, because it is the closest rival to W8 (skills that improve from real use) and it ships skill features every week, but it has three people, no named customer and does not run the work itself [site] [yc].

1. Fact file

2. Their proposition, reverse engineered

3. Positioning

• Category they claim: the shared context layer for "recursive companies" [yc].
• Category a YC partner would put them in: AI memory and context for AI-heavy teams, moving toward skills management (inference).
• What they want to be known for: the place where all of a team's AI work lives, so agents can reuse it and improve from it.
• Where they sit on the map: L1/L2/L3 = 3/1/1. Buyer: sme, business (individuals buy too).

4. Three layers vs ours

5. Same and different

6. Wedge overlap

7. Whitespace (where they stop)

• Layer gap: they record and measure; they do not run the agents or do the work [site] [jobs]. To own it we need one agent-team workflow run for 4+ weeks where accepted and rejected work changed the skills, with the rule-change log to show it.
• Buyer gap: they sell to AI-native builders first, with regulated and large buyers named as next [jobs]. Firms that use AI but are not AI-native (consultancies, agencies, accountancy firms) are not named anywhere (inference). To own it we need 3 to 5 such firms saying they have the pain.
• Number gap: no outcome number anywhere, only counts of imports and signups [blog]. To own it we need one before/after number, for example rework or failed runs falling over weeks.
• Price gap: none. Their prices are public and low ($28 to $80 a seat a month) [site], so a small buyer can pay.
• Speed gap: none. Download and a 7-day trial [site]. We are slower today.

8. For the YC application

• Proven today: Marton's logs show his system runs agents that do work and are checked by a different AI: about 5,600 jobs, about 160 failure records that each added an automatic check. On one workflow, offer feedback lines fell from 44 to 18 as 22 lint rules were added. Nessie's public material shows recording and skill measurement, not agents doing the work [site].
• Draft answer to "how are you different from Nessie?": DRAFT. "Nessie remembers what your agents did and tracks how your skills are used. Our system runs the agents, has a second AI check the work, and turns each failure into a rule, every day, on our own business." Founders say it their own way.
• Future claim (needs proof): "Skills that get better from every accepted or rejected piece of work, for a team that is not AI-native, with one number." Needs one outside team using it for 4+ weeks with a log of rule changes tied to checked runs, and a falling rework count.
• What we understand that they don't: a skill improves faster from checked work (accepted or rejected) than from reading traces. needs proof on a customer; Marton's logs show an early version on his own work (the 44 to 18 offer feedback figure).
• Copyability (M3): easy. Skill editing, analytics and "evaluation details" are already shipping in their changelog, and their stated plan is skills as the unit of improvement [site] [jobs]. Adding "propose edits from failed runs" is a feature for them, and it fits their buyer (inference). Running the agents themselves would pull them off a record-keeping product, so that part is closer to medium. The edge we would claim at W8 is not hard, so M3 can not score above 3 against Nessie.
• Why now (M4): None found in this teardown. The brief lists an August 2026 study of skill drift, but it helps Nessie as much as us, and I did not verify it here.
• Do not claim: "first to improve skills from real use" (Nessie claims the loop and ships skill analytics and evals) [yc] [site]. "Self-improving skills" as a track record (Marton's logs show anneal loops with only 1 to 4 weekly runs each). Any outside customer. "Company brain" or "shared memory" as a headline (Nessie owns those words).

9. Ideas worth taking

1. Skill analytics per skill: usage, success rate and a link to the runs behind each number [docs]. Makes "is this skill working?" a number, not a feeling.
2. Import the skills a user already has on day one [site]. First value without setup.
3. Review every changed skill file side by side before it installs [site]. Reviewable edits make people trust automatic changes.

10. Unknowns

• Whether the skill-improvement loop ships. Read the Skill evals and Skill analytics docs pages (the evals URL I tried returned 404) and try the Team plan trial.
• Real usage and paying teams. Ask in the Discord, or look for case studies after their next blog post.
• Round size and date. Look for a primary announcement or a Form D.
• Whether they sell in the UK. The site shows no UK customer, office or pricing.
• How the "Trusted by" logos use Nessie (one person each, or a team). Ask a user.

Sources

• nessielabs.com [site]
• Pricing [site]
• Changelog [site]
• Vision [site]
• Docs index and Skills docs [docs]
• Blog index [blog]
• The Recursive Company, 7 Jul 2026 [blog]
• Why We Killed Our Product Twice, 19 Apr 2026 [blog]
• YC company page [yc]
• Founding Forward Deployed Engineer job ad [jobs]
• Founding Engineer job ad [jobs]
• CB Insights: Nessie Labs and VCBacked: Nessie Labs [press] (funding totals; not a primary source)