XMarts room

Summer 2026

Agnost AI

We find where AI agents fail, then train better models to run them.

Visit website

What they do

reads agent conversations to find silent failures, then trains better models

How close to us

4, very close

direct W2: finds where agents silently fail and measures it, then trains models; buyer size not stated.

Wedges

  • W2 Seeing whether your AI agents actually produce direct
  • W8 The other 90% of an AI skill adjacent
L1 Company brain (knows the organisation) 1, light
L2 Agent team (runs agents, tools, approvals) 1, light
L3 Does and learns (finishes, checks, measures, updates the brain) 2, clear

Who buys

Firm size
sme
Buyer type
developer
Role
head of AI product, CTO or founder of a company that ships a customer-facing AI agent
UK
not known
Their own description (from YC)

Agnost AI reads every conversation a company's AI Agent has with its users to find where it's silently failing and then turns that same data into custom models that run the agent better, faster & cheaper than frontier models.

Full teardown

Agnost AI (YC Summer 2026): teardown vs XMarts

In one line: Agnost reads the conversations a company's AI agent has with its users, finds where the agent quietly fails, opens a pull request with a fix, and now also trains a cheaper custom model from the same data [site] [yc]. Threat: 2 because it is the same problem as W2 (see what your agents really produce), but it is two people, about $250k raised (reported), and it looks at chat with users, not at agents doing work [yc] [press].

1. Fact file

Item Finding Source
What they say Home page: "Your Agents, Self-Improving." YC line: "We find where AI agents fail, then train better models to run them." [site] [yc]
Founded, team, funding Founded 2025, San Francisco, 2 people, YC S26. About $250k pre-seed reported, Entrepreneurs First named as investor. The lead is not confirmed by a second source. Inc42 also lists Transpose Platform Management as a backer. [yc] [press]
Founders Shubham Palriwala (CEO): youngest engineer on Cisco's analytics team, first hire at Formbricks (open-source survey tool). Parth Ajmera (CTO): IIT Madras computer science, built data pipelines at Microsoft, led graphics engineering at Infurnia. What YC likely saw: two engineers from the analytics and data world, building the tool they would want themselves (inference from backgrounds). One aggregator names a different co-founder; YC and Inc42 do not, so treat that as an error. [yc] [press]
Customers They claim teams at Google, Exa and Corgi (YC page, earlier launch) and Exa and Comp AI (Inc42). The lists differ and no logo, case study or number appears on their site. Nature of the work with Google is not given. [yc] [press] [site]
Price Public. Free (1,000 events a month, 7 days kept), Starter $49, Pro $499 (1,000,000 events), Enterprise custom (VPC hosting, audit logs, fine-tuning). Job ads: none listed. [site] [yc]
Path Blog topics move in phases (inference from post dates): MCP server guides (Oct to Dec 2025), then "agent experience" metrics (Mar 2026), then conversation analytics and self-improving agents (May to Jul 2026), then fix pull requests and custom models. One post on MCP analytics was retitled but kept its old web address. No formal rename found. [blog]

2. Their proposition, reverse engineered

Part What they do
Who buys Teams that have shipped an AI agent to users: founders, product leaders, CTOs [blog]. Self-serve entry at $0 to $499 a month [site]. Enterprise by "Let's Chat" [site].
The pain A request returns "200 OK" and the tool call succeeds, yet the conversation went wrong, and nobody sees it unless they read the chat [press]. Their line: "You're measuring the pipe, not the water." [blog] Evals pass and users still leave [blog].
The promise Find silent failures, test a fix on past runs, and open a pull request for your team to review [site]. Later: a model trained on your own traces that is faster and cheaper than a frontier model [yc].
How it works Install by one prompt that adds an Agnost skill to Claude, Codex or Cursor, or by three lines of code or OpenTelemetry [site] [press]. It reads production chats and voice calls, groups them into patterns, traces each failure to a span, replays a suggested fix against past runs, then opens a PR [site]. A Slack agent answers questions about failures. Claude, Codex and Cursor can query it over MCP [site].
The output A list of failures with the share of runs affected, a trace of each, a tested pull request against prompts, tools or harness, and for enterprise a fine-tuned model [site] [yc].
Proof style Failure rates before and after on replay (their example: 4.5% to 0.4% repeated searches, labelled "illustrative") [site]. One real number set: a custom model for one unnamed customer's task, 71.5% to 87.9% task success, cost per 1,000 tasks $42.00 to $2.30, tested on 780 held-out traces. They say it is one model for one workload [yc]. They also claim over one million messages analysed a day, from a Product Hunt listing seen only through a search summary (UNVERIFIED).
Who sells it Founder-led. "Book a demo" and a founder call link; self-serve app; docs for 12 frameworks [site]. 129 blog posts, many written to catch search traffic [blog].
Price model Per event, in tiers, with retention as the lever. Founders Slack channel at Pro [site].
The story Enemy: dashboards, evals and logs that report success while users give up [blog]. Belief: agents should improve from real production conversations, and "automate the work, gate the decision": a person reviews and merges [blog].

3. Positioning

  • Category they claim: continuous improvement for AI agents; agent analytics; self-improving agents [site] [blog].
  • Category a YC partner would put them in: observability and evals for AI agents, with a fine-tuning upsell (inference).
  • What they want to be known for: the tool that turns real user conversations into shipped fixes for your agent.
  • Where they sit on the map: layers L1/L2/L3 = 1/1/2, buyer = sme, developer.

4. Three layers vs ours

Layer Them (score + evidence) Us (from the system inventory, "Marton's logs show")
L1 Company brain 1. Uses conversation data and a growing regression test set for one job: improving the customer's agent [site] [press]. No lasting memory of the company that other work reads. A vault of about 1,400 pages, 112 call transcripts and about 8,700 customer quotes, searchable by meaning.
L2 Agent orchestration 1. A fix-proposing agent and a Slack investigation agent, with a human approving each PR [site] [blog]. They do not run a team of agents; they watch someone else's. About 5,600 jobs run by agents from three makers, each checked by a different AI, then automatic tests.
L3 Execution and learning 2. Does the fix (PR), replay-tests it on past runs and compares failure rates [site]. Fine-tuning on traces is a route to level 3, but only one customer model is claimed [yc]. The result is checked, but a human merges it. Every failure gets a written record and an automatic check (about 160 records). 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 report success and silently fail; the failure is invisible in normal dashboards [blog]. Marton's logs show named failure classes (silent success, silent burn, quiet fleet) with a record and a check for each.
Fix the cause, not the symptom Same A fix is tested on past runs and added to a regression set [press]. Marton's logs show every incident must name a ratchet, an automatic check that stops it returning.
Human decides Same "Gate the decision": a person reviews and merges [blog]. Marton's logs show approvals sit with Marton on a review board.
What is watched Different Conversations between users and a product agent (chat, voice) [site]. The agents that do the work inside the business: jobs, merges, stuck queues, spend.
Buyer Different A team that ships an agent to its own customers [blog]. Target: SMEs and scale-ups that run agents to get their own work done (brief). No customer yet.
Owner of the result Different The customer's agent gets better; Agnost reports failure rates [site]. Not set yet. W2 would report whether work got produced.
How it gets in Different One prompt installs a skill; free tier [site]. The system is built for one person on three Macs; nobody else has run it.
Size Different 2 people, YC S26, $250k reported [yc] [press]. 2 founders, one builds.
Proof Different One anonymous model result; "illustrative" examples; names without cases [yc] [site]. Proof is internal logs; no outside number.
Price Different $0 to $499 a month, public [site]. Not set yet.

6. Wedge overlap

Wedge Fit Why
W2 Seeing whether your AI agents actually produce direct Same words: silent failures, what is stuck, a cost angle [site]. A partner would call them our competitor. The difference is the object watched: user chats with a product agent, not agents doing work.
W8 The other 90% of an AI skill adjacent They sell upkeep of an agent from real use: detect drift, test fixes, retire bad behaviour [blog]. It targets prompts and tools, not company skills. Their blog also argues agents stop improving after launch [blog]. Closest funded rival on this idea remains Nessie (brief).

7. Whitespace (where they stop)

  • Layer gap: they watch and fix one agent. They do not hold a company brain or run a team of agents [site]. To own it we need one customer's agent team watched and fixed from the company's own rules and data, with a before/after number.
  • Buyer gap: their view is the chat between a user and a product agent. Teams that run agents to do internal work (coding, back office) have no conversation to read, only jobs, merges and spend. Their blog touches coding agents but as a product feature [blog]. To own it we need a second team running agent work, who say "we cannot see what is stuck" and let us watch it.
  • Number gap: their one real number is for one anonymous model; the rest is labelled illustrative [yc] [site]. No customer money result is public. To own it we need a number a finance person would sign, such as hours of agent work lost to silent failures, found and cut over 4 weeks.
  • Price gap: none. They start free and top out at $499 a month [site]. We cannot beat that on price.
  • Speed gap: none. One-prompt install [site]. We cannot beat that on setup speed.

8. For the YC application

  • Proven today: Marton's logs show his system watches itself: a heartbeat reports what is stuck, what silently stopped and what it costs. About 160 failure records exist, each tied to an automatic check so it cannot return unseen. Agnost reads user chats with a product agent [site]; our evidence is about agents doing work inside a business.
  • Draft answer to "how are you different from Agnost AI?": DRAFT. "Agnost reads what users say to a company's customer-facing agent. We watch the agents that do the work inside a company, and we catch what is stuck, what stopped without a sound and what it costs." Founders say it their own way.
  • Future claim (needs proof): "We catch silent failures in agent teams doing real work and stop each one coming back." Needs one other team's agent work watched for 4 weeks, with a count of silent failures found and the share closed with a check.
  • What we understand that they don't: failures of agents doing work (stuck queues, silent stops, runaway spend) are a different shape from failures of chat, and five months of running such a fleet gave us named classes for them. needs proof outside Marton's own system.
  • Copyability (M3): medium. They could add job and spend views, since they already trace spans, read agent runs, and serve Claude Code and Codex over MCP [site]. But their data, their buyer and their pricing are built around user conversations, and they are 2 people with an enterprise and fine-tuning push [yc] [site] (inference). Not hard, so M3 can not score above 3 against Agnost alone.
  • Why now (M4): None found in this teardown. Do not use "agents fail silently now that they run more": that is general AI progress.
  • Do not claim: "we found agent observability" or "nobody watches agent fleets" (Agnost, and the Nessie and Wato rivals in the brief, do). "We have an observability product" (the inventory says it is not decoupled from our own queue). Any customer or revenue number for W2.

9. Ideas worth taking

  1. Install by one prompt that adds a skill to the user's own coding agent [site]. Distribution that fits people who already live in Claude or Codex.
  2. Show every fix as a replay on past runs with a failure rate before and after [site]. It turns "trust us" into a number.
  3. Plain names for failure types, written as a public list ("12 production failures your evals miss") [blog]. Our named failure classes could be shared the same way.

10. Unknowns

  • What Google, Exa, Comp AI and Corgi actually pay for or use. The lists differ by source. Ask a user, or look for a named case study later.
  • Real traction: events processed, paying accounts, revenue. None is public except an unconfirmed million-messages-a-day line.
  • Whether the model fine-tuning is the new focus or a side bet. The YC page leads with it; the home page leads with fix PRs.
  • The exact seed round, if any, and the lead investor. Check Entrepreneurs First and Crunchbase-type news later.
  • Whether they sell outside the US. Check their trust centre and docs for EU or UK hosting.

Sources

Technical detail
Slug
agnost-ai
Lens hash
179096c479
Pass
deep
Stale
no