Lark (YC Fall 2026): teardown vs XMarts

In one line: Lark sells five ready-made AI agents (buying, forecasting, stock watching, sales inbox, data questions) that sit on top of a wholesale distributor's ERP, and it says they can run a distributor's whole operation [site] [yc]. Threat: 2, because it is the same "one layer for one type of firm" idea as W5 with an early, 2-person team, no named customer and no result number [site] [yc]. Zero threat to W1 to W4.

1. Fact file

Item	Finding	Source
What they say	"Specialized AI Agents for Wholesale Distributors." YC line: agents "work with existing ERP systems, learning from each distributor's historical data" and team decisions.	[site] [yc]
Founded, team, funding	Founded 2026 (Dealroom says September 2026). 2 people. San Francisco. Dealroom lists $125K raised, which matches the usual YC deal (inference). Legal name LarkAI Technologies, Inc. Dealroom lists the site as trylark.ai, not larkagents.com (UNVERIFIED which is current).	[yc] [press] [site]
Founders	Jason Wang (CEO): they claim a YC-backed company, Datost, that grew ARR to $400K in 8 weeks; earlier built agent systems for industrial supply chain at Traba and worked at Confluent. Michael Wang: NYU Stern dropout, former national-level badminton player. YC likely saw a founder who has sold B2B AI fast and knows the warehouse and supply chain world (inference). Sources conflict on Datost: the YC page for its later form (E1O) lists one other founder and credits the $400K ARR to a founder bio; a search summary said Jason Wang is a co-founder (UNVERIFIED) [yc-datost].	[yc] [press]
Customers	None named. Site shows a demo workspace ("Acme Distribution") only. No case studies.	[site]
Price	Not public. No job ads (0 open roles), so no hint from hiring.	[site] [yc]
Path	No renames or pivots found. Company is weeks old.	[yc] [press]

2. Their proposition, reverse engineered

Part	What they do
Who buys	Wholesale distributors. The nine ERPs named (Prophet 21, Eclipse, Infor CSD, NetSuite, Acumatica, SAP Business One, Business Central, Sage 100/300, Spire, DDI) point to mid-sized distributors (inference). Buyer is a purchasing or operations head, or the owner (inference) [site].
The pain	Manual work in quotes, purchasing, stock and sales email. No buyer quote on the site [site] [yc].
The promise	Agents sit "on top of your ERP", run within limits you set, and expand to more workflows [site]. YC text says they can "run a distributor's entire operation" [yc].
How it works	1. Book a 30-minute call. 2. Connect ERP, email, files, accounting, CRM. 3. Pilot: map items, vendors and rules with the team and set up a first approved workflow. 4. Turn on more agents. Setup "around one to two months". [site] [docs]
The output	A ranked buying worklist, suggested order quantities, one draft purchase order and vendor email per supplier, an exceptions list for stock, CRM stage updates, plain-English answers with the numbers behind them. Orders under a set limit (example: $5,000, approved vendors) go through on their own; bigger ones wait for a buyer. An action log shows every read and write. [site] [docs]
Proof style	None yet. The site's numbers are sample dashboard figures (for example "$5.5k estimated sales at risk"), not customer results [site] [docs].
Who sells it	Founder-led. One path: book a meeting [site].
Price model	Not public [site].
The story	Distributors run on spreadsheets, email and a dated ERP; replace nothing, add agents on top (inference from the site's wording) [site]. Enemy: manual back-office work [yc].

3. Positioning

Category they claim: "Specialized AI Agents for Wholesale Distributors" [site].
Category a YC partner would put them in: vertical AI ops for one industry (inference). Same bucket as Whitespace (YC S26, "AI operating system for wholesale distributors") [yc-whitespace].
What they want to be known for: the AI back office for distributors.
Where they sit on the map: layers L1/L2/L3 = 1/2/1, buyer = mid-market (inference), business.

4. Three layers vs ours

Layer	Them (score + evidence)	Us (Marton's logs show)
L1 Company brain	1. The pilot maps items, vendors and rules for each customer, and every agent reads them [site]. The YC text says agents learn from history and decisions [yc], but the agent pages do not describe any learning from buyer decisions or overrides [docs]. Scored on the site, not the claim.	A vault of about 1,400 pages, 112 call transcripts and about 8,700 customer quotes, searchable by meaning. Learns from calls and checked work.
L2 Agent orchestration	2. Five named agents sharing one set of limits, approvals, pause and take-over, and a log of every action [site] [docs].	About 5,600 jobs run by agents from three makers, each checked by a different AI, then automatic tests. Built for one business on Macs.
L3 Execution and learning	1. Does the work (drafts and enters POs, updates CRM) [site]. No outcome check or measure anywhere: no stockouts-avoided, no savings, no accuracy [docs]. Human approval is the only check.	Every failure gets a record and an automatic check (about 160). Offer feedback lines fell from 44 to 18 as rules were added. Only about 15% of jobs reached the main code, and merges were unreliable in October.

5. Same and different

Topic	Same or different	Them	Us
Core belief	Same	Agents need the firm's own data and rules, and humans keep control [site].	Same (brief).
Approvals and limits	Same	Auto-run under a limit, wait above it, pause or take over any time [site].	Marton's logs show review by a different AI and automatic gates before work joins the main code.
Buyer	Different	Wholesale distributors with an ERP [site].	Not set. SMEs and scale-ups; one firm type is open (brief).
Kind of work	Different	Operational: buying, stock, sales inbox [site].	Coding agents and content, run for Marton's own business (system inventory).
Owner of the result	Different	The customer's buyers approve each action; no result is claimed [site].	Target: we own one number (brief). No customer yet.
How it learns	Different	Claimed in the YC text [yc]; not shown on the agent pages [docs]. Customer data is not used to train models [site].	Marton's logs show rules added from failures and feedback.
Size and speed	Different	Setup "one to two months" [site].	Target: value in days (brief). Unproven.
Delivery	Different	Software plus a 30-minute call and a pilot [site].	The system plus the two founders.
Proof	Same (both empty)	No customer, no number [site].	No outside customer pays yet (system inventory).
Price	Different	Not public [site].	Not set.

6. Wedge overlap

Wedge	Fit	Why
W5 One layer for one type of small firm	direct	Same shape: one operating layer on top of a firm's existing system, for one firm type, built once and sold to many [site]. Theirs is distributors; our example is accountancy firms (brief). They are first to the pattern in this batch; Whitespace (S26) is on the same buyer [yc-whitespace].
W7 Find and fix margin leaks	adjacent	Their Data analysis and Inventory agents find overstock, stockout risk and stalled deals [site]. They stop at flags and drafts. They do not claim margin work or a before/after number [docs].

7. Whitespace (where they stop)

Layer gap: they do the work but do not check or measure the result, and do not show learning from outcomes [docs]. To own it we need a run where the agents' output is checked and a rule changes because of it, with rework falling.
Buyer gap: the product is tied to nine named ERPs and US hosting [site], so it skips firms without an ERP and non-US firms (inference). To own it we need buyers outside distribution saying they have the pain.
Number gap: no customer result or case study [site]. To own it we need one workflow with a before/after number a finance person would sign.
Speed gap: "one to two months" to set up [site]. To own it we need first value in under 14 days.
(Price gap not asserted: their price is not public.)

8. For the YC application

Proven today: Marton's logs show his system already runs agents that do work and have it checked by a different AI before it counts: about 5,600 jobs and about 160 failure records, each adding an automatic check. Lark's public pages show actions with approval limits but no checked result or number [site] [docs].
Draft answer to "how are you different from Lark?": DRAFT. "Lark gives distributors ready-made agents for one industry. Our system already does and checks work every day on our own business, and we want to prove one number on one workflow." Founders say it their own way.
Future claim (needs proof): "Our agents get better from checked results, and we show the number." Needs one customer workflow run 4+ weeks with a before/after number, and a rule-change log tied to those runs.
What we understand that they don't: the value is in the checked result feeding the next run, not in the agent list. needs proof on a customer; Marton's logs show it on his own work (offer feedback lines fell from 44 to 18).
Copyability (M3): easy. They already claim learning from decisions [yc], so adding result checks and a measure is a feature for them, probably weeks. Their real moat is ERP connectors and the distributor buyer, which does not help against a rival in another firm type (inference). Not hard, so M3 can not score above 3 against Lark.
Why now (M4): None found.
Do not claim: "we are the AI OS for X" (Lark and Whitespace use that frame [yc] [yc-whitespace]). "Our agents learn from outcomes" for any customer (no customer run). Anything about distributors being a market we serve.

Pitch note: Lark, Whitespace, Didero and Faction all sell AI agents to distributors or their buyers (see Sources). A YC partner will see "AI layer for one firm type" as a funded pattern. YC took a 2-person team with no named customer in Fall 2026 [yc], so the bar is the founders and the market. Our edge must be the loop plus one number, not the vertical.

9. Ideas worth taking

Show limits as a first-class control: "orders under $X go through, bigger ones wait" [site]. Buyers see where the human stays.
Name the agents by job (Purchasing, Inventory) and list the ERPs on the page. It reads as concrete, not as "AI platform".
A log showing every read and action, with system and time [site]. It answers the data-safety question up front.

10. Unknowns

Whether any customer or pilot exists. Look for their YC launch post or a demo-day page.
Real price and contract size. Ask a distributor who has seen a quote.
Whether the "learns from decisions" claim is built. Ask for a demo of an override changing the next suggestion.
Whether trylark.ai (Dealroom) or larkagents.com is the live site.
Faction's raise and date: the MDM article page returned an error, so only the headline ("Faction Raises $4M") was seen.
Whether Datost's $400K ARR claim is true: unverified; the only source is the founder's own YC bio.

Sources

larkagents.com [site]
Purchasing agent page [docs]
Inventory monitoring agent page [docs]
Lark FAQ [site]
YC company page: Lark [yc]
Dealroom: Lark [press]
YC company page: datost slug (now E1O) [yc-datost]
YC company page: Whitespace [yc-whitespace]
Didero $30M Series A, Digital Commerce 360, Feb 2026 [press]
Faction raises $4M, MDM (headline only; page not readable) [press]

Lark

What they do

How close to us

Wedges

Who buys

Full teardown