Agnost AI reads every conversation your agent has - then files the fix

The idea for Agnost AI arrived the way most good ideas do: after something failed. Shubham Palriwala and Parth Ajmera, childhood friends of more than eight years, had built an AI product and were watching it not work. So they opened up their Discord to explain to the team what had gone wrong. Somewhere in the middle of typing out the post-mortem, they noticed the obvious thing. The answer was already there. It was sitting in the transcripts of conversations their product had been having with real users the whole time. Nobody had been reading them.

That is the small, human observation Agnost AI is built on, and it turns out to be a big one. Software that talks to people generates a river of feedback every hour of every day. When your product is a chatbot or a voice agent, every session is a user telling you - in plain language - what they wanted, where they got stuck, and why they gave up. Most teams never look. There is too much of it, and reading logs is nobody's favorite job.

Agnost AI, a company in Y Combinator's Summer 2026 batch, does the reading. It ingests the conversations your chat or voice agent has in production, clusters them into product-specific intents - feature request, bug, frustration, churn risk - and then does the part no dashboard does. It opens a pull request against your prompts, tools, and harness to fix what it found.

01 / THE PROBLEMYour evals pass. Your users still leave.

Teams building agents lean on evals - scripted tests that check whether the model does what you expect. The trouble is that evals only test what you thought to test. Real users are more creative than any test suite. They phrase things sideways, ask for features that don't exist, misunderstand the agent, and rage-quit in ways no engineer anticipated. The agent passes every test and quietly loses the customer anyway.

Agnost's tagline names the gap directly: catch agent failures your evals miss. Instead of grading the agent against a fixed rubric, it reads what actually happened and lets the failures name themselves.

Give us your traces and we'll personally find 3 insane things going on in your agent that you'd have no idea about. The founders' standing offer to prospects

02 / HOW IT WORKSFrom transcript to merged pull request

Installation is deliberately boring, which is the point. Three lines of code, or you point an existing OpenTelemetry exporter at Agnost and you are done in about two minutes. It is model-agnostic and framework-agnostic - hence the name - so it works with whatever stack you already run, with custom metadata attached.

The last step is the one that raises eyebrows. Agnost's GitHub app doesn't just flag a problem; it drafts a change, checks it against past conversations to see whether it would have helped, and files a pull request. A human still reviews and merges. But the machine has done the reading, the diagnosis, and the first draft of the repair.

What Agnost pulls out of raw conversations

Feature reqs

high volume

Bugs

reproducible

Frustration

sentiment

Churn risk

early signal

SOP breaks

compliance

03 / WHO USES ITA million events a day, and counting

Agnost reports processing more than a million events every day. Its named customers span the range from frontier labs to insurance startups: Google, Exa, Corgi Insurance, Orchid, Lopus AI, Odysser, and Comp AI. The results the company points to are unusually concrete for an early-stage tool. Lopus AI merged 16 of 18 pull requests that Agnost's system opened on its own. Odysser surfaced 1,247 feature requests that had been buried in user chats. Corgi's voice BDRs booked more meetings after the fixes went in.

1,247 feature requests were sitting in the chat logs. The team hadn't read them. The software had. On Odysser's early results with Agnost

04 / THE BUSINESSPriced by the event, free to start

The model is straightforward usage-based SaaS, metered on monthly events. A free tier lets teams try it on up to 1,000 events a month with a week of retention; from there it climbs to a paid Pro plan sized for a million events, and an Enterprise option that adds self-hosting for teams that can't send conversations to someone else's cloud.

05 / THE FOUNDERSCisco's youngest, and IIT Madras rank 159

Shubham Palriwala, the CEO, was the youngest engineer on Cisco's analytics team and the first hire at Formbricks, with open-source contributions to Bitcoin, OWASP, and the Linux Foundation. Parth Ajmera, the CTO, studied computer science at IIT Madras - where he ranked 159 out of roughly 1.5 million candidates - built terabyte-scale Spark pipelines at Microsoft, and led graphics engineering at Infurnia. They have been building things together for the better part of a decade.

In April 2026, the company was reported to have raised a pre-seed round of around $250,000, with Entrepreneurs First named as lead investor. The team is small - four people - and split between San Francisco, where the company is registered, and an engineering base in Bengaluru.

06 / THE MARKETAnalytics for software that talks back

Agnost sits in the fast-filling category of LLM observability and agent analytics, alongside tools like LangSmith, Langfuse, Braintrust, and Arize Phoenix. Most of those focus on tracing and scoring - showing you what the model did and how it graded. Agnost's wager is different: that the valuable signal lives in the full conversation, read for product intent, and that the useful output isn't another dashboard but a fix. Reading is the commodity; shipping the pull request is the product.

Your users already told you what's wrong. You just weren't listening at scale. The premise behind Agnost AI

Whether autonomous, self-filing pull requests become normal or stay a curiosity is an open question. Letting a tool open changes against your codebase asks for a level of trust the industry is still negotiating. Agnost's answer is to keep a human on the merge button and to test every proposed change against what really happened before anyone sees it. For teams drowning in conversations they'll never have time to read, that trade looks worth making.