Moda (YC Winter 2026): teardown vs XMarts

In one line: Moda reads every trace from a team's production AI agent, groups the silent failures into ranked problems, and tests a fix against the failing traces before it ships [site] [docs]. Threat: 2 because it is the same problem as W2 and it is live, but it watches agents that talk to end users, it does not do the work, and it names only one customer [site] [blog].

1. Fact file

Item	Finding	Source
What they say	"Your agents should never make the same mistake twice." YC launch line: "Turn production agent traces into validated improvements for your agent harness."	[site] [yc]
Founded, team, funding	Founded 2025, YC W26, 6 people, San Francisco. Funding UNVERIFIED: aggregator snippets say $500k from the accelerator, and one says an undisclosed seed from Pioneer Fund in March 2026. No primary announcement found.	[yc] [press]
Founders	Pranav Bedi (CEO): data infrastructure at banks and insurers, then AI agents at Shopify and Cerebral Valley. Mohammed (also spelled Mohammad) Al-Rasheed (CTO): AI agents and infrastructure at Shopify, Notion and Clio. Both left Waterloo. What YC likely saw: two engineers who had run production agents at known firms (inference).	[yc] [blog]
Customers	One named case study: Octolane, an AI CRM. They claim 30+ customers and 50M conversations processed, with no other names. The customers page lists none.	[blog] [site]
Price	Custom, "scaled to production conversation volume and pipeline usage". Access is invite-only. No job ads (YC page shows 0 open roles), so no hint of sales cost.	[site] [docs] [yc]
Path	Launch post framed it as "The Continual Learning Layer for AI Agents" (YC post, about Feb 2026, inference from "8 months ago" and W26). Site title became "Production monitoring and evals for AI agents". Early-access pages for 2026 Q1 and Q2 use "harness engineering" and "Autonomous Harness Engineering". Brand refresh on 11 Oct 2026.	[yc] [site] [docs]

2. Their proposition, reverse engineered

Part	What they do
Who buys	Teams that run an AI agent in production for users: engineers and AI product leads [site]. Octolane (a small AI CRM) is the one named buyer [blog]. Size is (inference) sme from that one case.
The pain	Agents fail without an error: frustrated users, failing tools, lazy answers, false "done" claims, hallucinations. Teams learn from complaints and read transcripts by hand [site] [yc].
The promise	Find each repeated mistake, understand the cause, and know whether the fix worked, so the same mistake does not come back [site].
How it works	Install an SDK or send OpenTelemetry. Moda reads every trace, scores signals, groups traces into problems ranked by users hit, and replays a proposed prompt, model or tool change against the failing traces [docs] [site]. Humans make the change and decide what ships [site].
The output	A ranked problem list with evidence, alerts in Slack, Linear or email, Linear issues, a before/after check on the fix, and a "hand this to Cursor, Claude Code or Codex" button [site] [blog] [docs].
Proof style	Counts of scale (50M conversations, 150M segments) and one case study (43,000 chat sessions, 5 shipped fixes). No money or rate result. They say 78% of cluster labels were rated actionable by reviewers (their test) [blog].
Who sells it	Founder-led. "Book a demo" on your own data within 60 minutes. Invite-only sign-up [blog] [docs].
Price model	Custom, by conversation volume [site].
The story	Everyone has the same models, so the teams that learn fastest from real usage win (as reported by a third party, not from their site) [press]. Enemy: hand-reading logs and curated test sets that miss real traffic.

3. Positioning

Category they claim: "Production monitoring and evals for AI agents" and "harness engineering" [site] [docs].
Category a YC partner would put them in: observability for AI agents, the "Sentry for agents" slot (inference; Moda has its own page using that phrase [site]).
What they want to be known for: catching silent agent failures and closing each one with proof.
Where they sit on the map: layers L1/L2/L3 = 1/0/1, buyer = sme (inference), developer.

4. Three layers vs ours

Layer	Them (score + evidence)	Us (Marton's logs show)
L1 Company brain	1. Builds a use-case map of what users ask, user cohorts, and a map of the customer's agent code (`moda init`) [docs] [blog]. It is memory of the agent's usage, not of the organisation.	A vault of about 1,400 pages, 112 call transcripts and about 8,700 customer quotes, searchable by meaning.
L2 Agent orchestration	0. Does not run agents. It hands a finding to Cursor, Claude Code or Codex, and sends nothing until the user presses Enter [docs]. A replay sandbox runs the customer's own agent [docs].	About 5,600 jobs run by agents from three makers, each checked by a different AI, then automatic tests.
L3 Execution and learning	1. Checks only. It replays changes, then after a Linear issue closes it compares the signal rate before and after, waiting at least 7 days [docs] [blog]. A fixed problem that recurs is reopened [docs]. Engineers make the change [site].	About 160 failure records, each adding an automatic check. Offer feedback lines fell from 44 to 18 as rules were added. Only about 15% of jobs reached main, and merges were unreliable in October.

5. Same and different

Topic	Same or different	Them	Us
Core belief	Same	Agents fail silently and need a record of each failure [site].	Marton's logs show every failure gets a written record and an automatic check.
Kind of failure watched	Different	Conversations between users and an agent: frustration, bad answers, tool errors [site].	Marton's logs show jobs, queues, merges and cost: what is stuck, what silently stopped, what it burns.
Buyer	Different	Teams with a customer-facing agent [site].	Target: SMEs and scale-ups (brief). Wedge W2 buyer: teams running coding agents in loops (inventory).
Who does the work	Different	Engineers make the fix; Moda tests it [site].	Agents do the work and a different AI checks it, on Marton's own business only.
How it learns	Similar	A fix is checked against later traces; recurring problems reopen [docs].	Failures become automatic checks (inventory). The check is rules, not a replay of past traffic.
Delivery	Different	SaaS, SDK, CLI, MCP server, invite-only [docs].	Mac and launchd bound, one user. Nobody else has run it (brief).
Proof	Different	Scale counts and one case study [blog].	No outside customer yet (brief).
Price	Unknown	Custom by volume [site].	Not set yet.
Speed	Similar	Demo on your own data in 60 minutes [blog].	Not tested.

6. Wedge overlap

Wedge	Fit	Why
W2 Seeing whether your AI agents actually produce	direct	Same promise: what silently failed and whether the fix worked [site]. They stop at conversations and traces. Stuck queues, cost and merge state are not in their public pages (UNVERIFIED that they have none). They also accept coding-agent traces [docs], so the gap may be small.
W8 The other 90% of an AI skill	adjacent	They name "skills" and "memory" among the things they improve [site], and the YC post says "validated improvements" [yc]. No public sign they retire or refresh skills. Their Q2 early-access page title ("Autonomous Harness Engineering") is a signal they may be heading toward self-improvement [site]. Close to Nessie, the closest rival in the brief.

7. Whitespace (where they stop)

Layer gap: they watch and test; engineers make the change [site]. To own it we need agents that make the change, a different AI that checks it, and a number that moves, on one customer workflow for 4+ weeks.
Buyer gap: they watch what end users say to an agent. They say little about teams whose agents work overnight on internal jobs (inference from the pages read). To own it we need teams running coding agents in loops to say the pain is real.
Number gap: the case study counts sessions and fixes, not a result a finance lead would sign [blog]. Their Linear check does compare a signal rate before and after [docs], so the gap is the money or time number. To own it we need one before/after number from a customer.
Price gap: not a gap we can claim. Their price is custom and not public [site].
Speed gap: not a gap. They offer a demo on your data in 60 minutes [blog].

8. For the YC application

Proven today: Marton's logs show his system watches itself: a heartbeat for what is stuck or silently stopped, a cost meter, and about 160 failure records that each added an automatic check. It does this for agents that do the work, not for chat. Moda's public results are about reading traces, with one named customer [blog].
Draft answer to "how are you different from Moda?": DRAFT. "Moda watches what your agent says to users and tests fixes. Our system watches agents that do the work overnight, what got stuck, what failed quietly, what it cost, and it has done that on our own business for five months." Founders say it their own way.
Future claim (needs proof): "We catch the failures of unattended agent work and close them with a check." Needs one outside team running coding agents on our setup for 4+ weeks, with a before/after number such as stuck jobs or cost per merged job.
What we understand that they don't: unattended agent work fails in ways a chat trace never shows (stuck queues, silent stops, merges that never land, runaway cost). needs proof outside Marton's own system. Marton's logs show the failure list, but it is one user.
Copyability (M3): medium. Trace-level checks for coding agents are a feature for them: they already take Claude Agent SDK and coding-agent traces and ship often (10 changelog entries between 2 and 12 Oct 2026) [docs] [site]. A job queue, merge and cost view is months of work and pulls them off conversation analysis (inference). Not hard, so M3 can not score above 3 against Moda.
Why now (M4): None found in this teardown. Do not use "agents are everywhere now": that is general AI progress.
Do not claim: "nobody monitors agents" (Moda and the rivals it names on its own comparison pages do). "Continual learning" or "self-improving" as a headline (Moda uses both words, and our anneal loops have 1 to 4 weeks of history). "We catch failures nobody sees" (no proof against Moda's coding-agent support). Anything about W2 being uncrowded.

9. Ideas worth taking

A fix is not closed until a later check says so: wait at least 7 days, compare the signal rate before and after, and reopen on recurrence [docs]. Same shape as our ratchet rule.
A "hand this to your coding agent" button that opens Claude Code, Codex or Cursor with the evidence prefilled and sends nothing until the user presses Enter [docs].
Public comparison pages against each rival, plus llms.txt and an agent instructions file, so AI search and AI agents can find and explain them [site]. They run about ten such pages.

10. Unknowns

Real funding and investors. Check Moda's own announcements or the YC Demo Day list. One WFTV article was blocked in our region and could not be read.
Who the 30+ customers are, and whether any run coding agents. Ask on the pitch page or look for logos after Demo Day.
Price. Ask a buyer who has had a quote.
Whether the Q2 "Autonomous Harness Engineering" page means they will generate the fix themselves. The page text did not load in this run.
Whether they sell in the UK. No UK signal found.
Security and compliance claims. The security page returned 404 when fetched, and the docs say little.

Sources

moda.dev [site]
Moda pricing page [site] (custom pricing, from the page's summary)
Moda customers page [site]
Moda llms.txt [site]
Moda changelog, 12 Oct 2026 [site]
Moda Weekly, issue 3, 5 Oct 2026 [blog]
Octolane case study, 14 May 2026 [blog]
Clustering Agent Failures from Traces, 24 Apr 2026, updated 2 Sep 2026 [blog]
Moda blog index [blog]
Moda docs [docs]
Docs: Fixes [docs]
Docs: Register a harness [docs]
YC company page [yc]
YC launch post [yc]
Respan market map: Moda [press] (a rival's page, so not neutral)
Neuronfeed / Caplight / Preqin profiles via search [press] (funding snippets, UNVERIFIED)

Moda

What they do

How close to us

Wedges

Who buys

Full teardown