Agnost AI launched this week with a sharp pitch from co-founder and CEO Shubham Palriwala: "with AI agents, there's no complaint, just churn." The product is positioned as the self-improving layer for any AI agent in production. It captures what users actually want, links those signals to where your agent gets it wrong, and ships fixes automatically.
What it does
Agnost wraps your agent and runs three loops in parallel.
- Intent capture. Every input and output is parsed for what the user wanted and how the agent responded. Beyond binary success or failure, it classifies intent and sentiment so you can see patterns of what is working and what is missing.
- Continuous evals. Production traffic is evaluated in real time against guardrails and quality criteria, instead of relying on weekly manual sampling or offline test sets.
- Autonomous improvement. Insights from intent and eval data feed back into the agent, closing the loop without a human rewriting prompts each cycle.
Setup is two minutes. OpenTelemetry native. SDKs in TypeScript and Go. Works with any LLM and any framework.
The dashboard includes a Spotlight feature for querying aggregated conversation data in natural language, and Real-Time Production Evaluations that flag guardrail violations before they hit users.
Where it sits in the stack
LLM observability tools like Langfuse, Helicone, and Braintrust focus on tracing, cost, and latency. They answer "what happened?" but not "did the user get what they came for?" Agnost is in a different lane, centered on intent and closed-loop improvement. The closest direct comparison today is Shinzo Labs.
Agnost's public writing on Intent Resolution Rate makes the underlying argument concrete. A user who rephrases the same question is telling you the first answer missed. A user who escalates to a human is telling you the agent failed. A user who copies the output is telling you it worked. These behavioral signals are what Agnost classifies and acts on.
Their Exa case study reports time from signal to shipped fix dropped from 14 hours to zero.
Team and backing
Co-founded by Shubham Palriwala (ex-Formbricks founding engineer, ex-Cisco) and Parth Ajmera. Entrepreneurs First S25 batch, $250k raised. Backed by Transpose Platform Management. Incorporated in Delaware as Agnost Tech Inc, with engineering presence in Bangalore and San Francisco.
Public customer logos: Exa, Google, Comp AI, Omelo, Wavelength, Hugeicons, Orchid, Clad Labs, Super Intern, Decawork.
Why it is worth watching
Most agent teams today measure token spend and latency because those numbers are easy to graph, and use thumbs up ratings or CSAT scores as a proxy for quality. Agnost is arguing that intent resolution is the actual quality signal, and trying to make it the new default. The category will not stay single-vendor for long, but Agnost is one of the first to ship around the idea, and the framing alone is worth absorbing.