Agnost AI is an observability and improvement layer for teams shipping LLM-powered agents in production. It extracts intent signals from every conversation, runs evals continuously, and ships improvements to your agents autonomously , so issues that would normally show up as silent drop-off (a user asking for something the agent fumbled, then never coming back) become structured signals that feed back into the agent. The pitch in the launch tweet is blunt about this dynamic. With AI agents, users rarely file tickets, they just stop using the product, which makes capturing raw user intent the missing input most teams lack today.
The product is positioned at the eval and agent-tuning layer of the stack, sitting between the agent runtime and whatever model or framework a team is already using. Agnost AI works with any LLM, any framework, with a 2-minute setup, and is OpenTelemetry native , which matters for teams who have already instrumented traces and want to layer intent extraction and continuous evaluation on top rather than swap toolchains. It captures every signal, runs evals continuously, and ships improvements without manual intervention , which is the part that distinguishes it from passive observability tools that surface problems but leave the fix to engineers.
The company is led by co-founder and CEO Shubham Palriwala, headquartered in Delaware with engineering and operational presence in Bangalore and San Francisco , and is backed by Entrepreneurs First . Palriwala was previously a founding engineer at Formbricks and worked on analytics at Cisco , a background in instrumentation and product telemetry that lines up directly with what Agnost is building. The launch is timely because most teams now have agents in front of users but no closed loop between user behavior and agent behavior, and Agnost is making a specific bet that the loop should be automated rather than left to weekly prompt reviews.
Comments (14)
Sign in to join the discussion.
P
Priya RaghunathanMay 1, 2026
"No complaint, just churn" is the kind of line that should be tattooed on every B2C founder's forearm. Genuinely the cleanest framing of the agent feedback problem I've seen this month.
J
Juno AdeyemiMay 1, 2026
Also followup, what happens when the "raw intent" the model infers is just wrong? Feels like you'd be improving the agent toward a hallucinated user need.
K
kostyaMay 1, 2026
The launch video cut at 0:08 where the failed agent response fades into the dashboard is doing actual work. Whoever edited this understood that you sell the pain before the product.
F
Finn O'HalloranMay 1, 2026
Curious what the wedge looks like when every model lab ships their own eval and improvement loop natively. Interesting space, watching the retention story.
J
Juno AdeyemiMay 1, 2026
Naive question but how does it know the agent did something wrong if the user never says anything? Like is it inferring from drop-off or is there a labeling step somewhere?
D
DeShawn K.May 1, 2026
The tweet copy buries the lede. "There's no complaint, just churn" should have been line one, the product name can wait.
E
Emeka O.May 1, 2026
The wordmark has way too much letter spacing for how short "Agnost" is. Tighten the kerning by like 4% and the whole landing page snaps together.
L
Latif the LateMay 1, 2026
So this is like... Mixpanel for agents? I feel like we figured out session replay in 2019 and now we're rediscovering it with extra steps.
M
Marisol QuintanaMay 1, 2026
How big is the team shipping this? Because the landing page, the video, AND the tweet thread all hitting at once usually means someone didn't sleep this week.
B
Ben ThonkarMay 1, 2026
Every agent eventually trains on its own confused users. The question is whether the loop converges or just gets confidently weirder.
K
Kat VolkovaMay 1, 2026
"Autonomously improves the agent" is doing a lot of heavy lifting here. I'll believe the demo when I see what it does on a workflow that isn't a happy-path support bot.
R
Rohit BhallaMay 1, 2026
First three seconds of the video are the founder talking to camera. That's a retention crime in 2025, lead with the broken agent and the fix, talking head goes at the end.
S
Sanna LindqvistMay 1, 2026
Cool, I'm 3 weeks out from shipping something adjacent so I'll be studying this rollout closely. Mostly to figure out how you got the EF tweet timing this clean.
T
Tahir MansourMay 1, 2026
We built something with this exact pitch internally in 2019 except we called it "intent reconciliation" and it died in a planning doc. Glad someone is actually shipping the idea outside a 200 person org.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Agnost AI
Your agents should get better every day.
Your agents should get better every day.
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Agnost AI is an observability and improvement layer for teams shipping LLM-powered agents in production. It extracts intent signals from every conversation, runs evals continuously, and ships improvements to your agents autonomously , so issues that would normally show up as silent drop-off (a user asking for something the agent fumbled, then never coming back) become structured signals that feed back into the agent. The pitch in the launch tweet is blunt about this dynamic. With AI agents, users rarely file tickets, they just stop using the product, which makes capturing raw user intent the missing input most teams lack today. The product is positioned at the eval and agent-tuning layer of the stack, sitting between the agent runtime and whatever model or framework a team is already using. Agnost AI works with any LLM, any framework, with a 2-minute setup, and is OpenTelemetry native , which matters for teams who have already instrumented traces and want to layer intent extraction and continuous evaluation on top rather than swap toolchains. It captures every signal, runs evals continuously, and ships improvements without manual intervention , which is the part that distinguishes it from passive observability tools that surface problems but leave the fix to engineers. The company is led by co-founder and CEO Shubham Palriwala, headquartered in Delaware with engineering and operational presence in Bangalore and San Francisco , and is backed by Entrepreneurs First . Palriwala was previously a founding engineer at Formbricks and worked on analytics at Cisco , a background in instrumentation and product telemetry that lines up directly with what Agnost is building. The launch is timely because most teams now have agents in front of users but no closed loop between user behavior and agent behavior, and Agnost is making a specific bet that the loop should be automated rather than left to weekly prompt reviews.
Comments (14)
"No complaint, just churn" is the kind of line that should be tattooed on every B2C founder's forearm. Genuinely the cleanest framing of the agent feedback problem I've seen this month.
Also followup, what happens when the "raw intent" the model infers is just wrong? Feels like you'd be improving the agent toward a hallucinated user need.
The launch video cut at 0:08 where the failed agent response fades into the dashboard is doing actual work. Whoever edited this understood that you sell the pain before the product.
Curious what the wedge looks like when every model lab ships their own eval and improvement loop natively. Interesting space, watching the retention story.
Naive question but how does it know the agent did something wrong if the user never says anything? Like is it inferring from drop-off or is there a labeling step somewhere?
The tweet copy buries the lede. "There's no complaint, just churn" should have been line one, the product name can wait.
The wordmark has way too much letter spacing for how short "Agnost" is. Tighten the kerning by like 4% and the whole landing page snaps together.
So this is like... Mixpanel for agents? I feel like we figured out session replay in 2019 and now we're rediscovering it with extra steps.
How big is the team shipping this? Because the landing page, the video, AND the tweet thread all hitting at once usually means someone didn't sleep this week.
Every agent eventually trains on its own confused users. The question is whether the loop converges or just gets confidently weirder.
"Autonomously improves the agent" is doing a lot of heavy lifting here. I'll believe the demo when I see what it does on a workflow that isn't a happy-path support bot.
First three seconds of the video are the founder talking to camera. That's a retention crime in 2025, lead with the broken agent and the fix, talking head goes at the end.
Cool, I'm 3 weeks out from shipping something adjacent so I'll be studying this rollout closely. Mostly to figure out how you got the EF tweet timing this clean.
We built something with this exact pitch internally in 2019 except we called it "intent reconciliation" and it died in a planning doc. Glad someone is actually shipping the idea outside a 200 person org.