The archive · Developer & Business Tools · Product decision · 2025–2026
Agnost AI bets agent conversations are the next feedback channel
YC S26 reads every user-agent conversation to surface frustration and feature demand; ~1M messages/day, $499/mo Pro, then trains specialist models from traces
Agnost AI
What the business is
Product analytics for conversational AI: reads the conversations an agent has with users, clusters them into intents and failures, and surfaces the highest-impact fixes with the underlying conversations and traces.
How it started
Childhood friends Shubham (Cisco analytics, first hire at Formbricks) and Parth (IIT Madras, Microsoft Spark, GPU lead at Infurnia) spent about a year building agent tooling; founders of AI assistants kept asking for a 'PostHog for conversations', so they built Agnost, joined YC S26 and launched in July 2026.
What happened
At launch they were live with multiple companies at ~1M messages/day; one AI video editor customer found around 70 users wanting auto-subtitles, phrased as 'add this text in this frame', 'can you caption it', 'give me a transcript' and variations across languages. A month later (2026-08-26 Launch YC) they extended the bet: train workload-specific models from the same traces — the first benchmark vs Opus 4.8 on 780 held-out traces showed task success 71.5%→87.9%, median latency 4.80s→0.47s, and cost $42.00→$2.30 per 1,000 tasks.
How it ended up
Still running: two founders, YC S26, active in San Francisco, expanding from conversation analytics into customer-trained specialist models as of September 2026.
Background
Agnost AI is a YC S26 startup founded by childhood friends Shubham and Parth, betting that teams shipping chat and voice agents will pay for product analytics built for language. Their argument: a trace can say 'success' while the user still got nothing useful, and users rarely file explicit feedback — they curse, rephrase, correct the agent, or leave.
The company reads production conversations through SDKs or OpenTelemetry, clusters them into product-specific intents, and surfaces failures like rage-prompting, repeated requests and hidden feature demand. At launch it was live with multiple companies, ingesting roughly 1 million messages a day, with public pricing from free to $499/month Pro and SOC 2 Type 1 compliance.
In August 2026 the founders extended the bet: the same traces became training data. Their first workload-specific model beat Opus 4.8 on 780 held-out customer traces — task success up 22.9% (71.5%→87.9%), median latency down 90.2% (4.80s→0.47s), cost per 1,000 tasks down 94.5% ($42→$2.30) — turning logs from debugging data into evaluation data and then training data.
What has to be true
- Chat and voice products lack the metrics of web apps: when the interface is language, funnels and clicks lose meaning, so technical observability answers the wrong question.
- Users give feedback implicitly — repeated rephrasing, corrections, all-caps rage — which keyword and SQL analysis miss and static evals cannot enumerate in advance.
- One customer found around 70 users wanting auto-subtitles expressed in a dozen different ways; that discovery loop is exactly what a PostHog-for-conversations product sells.
- The data gathered for analytics doubles as proprietary training data, letting the company move from $499/month software toward higher-value specialist-model contracts.
What can be applied
When the interface becomes language, clicks and funnels go quiet — build the measurement layer for that channel early, and the data becomes both product and training set.
Aftermath
As of 2026-09-02, Agnost AI is active with two founders in YC S26, based in San Francisco. It is live with multiple customers at roughly 1 million messages per day, publicly priced from free to $499/month, SOC 2 Type 1 compliant, and expanding from conversation analytics into customer-trained specialist models, with a first benchmark published 2026-08-26.
Sources
- Launch HN: Agnost AI (YC S26) – Extract user feedback from agent conversations
- Agnost AI: Product Analytics for Conversational Agents
- Agnost AI: Turn Your Agent Traces Into a Faster, Cheaper, & more Accurate Custom Model
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card