Agnost AI Reads the Rage-Prompts Your Evals Never See
Your eval suite says the agent answered correctly. The user was cursing at it two turns earlier and left without converting. Both things are true. Only one of them shows up on your dashboard.
Agnost AI, a YC S26 company that launched on Hacker News today, goes after exactly that blind spot. It reads production chat and voice conversations and pulls out behavioral failure signals: users rage-prompting (which is to say, swearing at your agent), rephrasing the same request five different ways, correcting the agent, asking for features that do not exist, or simply walking away after a response that was technically a success. Point it at your traffic and it generates the failure categories that actually apply to your product β broken workflows, retry loops, setup friction, churn risk β and then opens reviewed pull requests to fix them.
The framing is sharp because it names the gap the entire eval industry has been dancing around. Benchmarks measure whether the agent did the task. They cannot see the human on the other side deciding it is not worth the trouble. Task success and user success are different variables, and only one of them pays you.
The company came out of YC S26 with a 250 thousand dollar pre-seed backed by Entrepreneurs First, founded by Shubham Palriwala and Gourab Saha. The hard part is precision: frustration is a noisy signal, and a system that cries wolf on every rephrase becomes one more dashboard nobody opens. But the raw material is already sitting in every agent company's database, unread. Somebody was going to go mine it.
https://agnost.ai/
← Back to all articles
Agnost AI, a YC S26 company that launched on Hacker News today, goes after exactly that blind spot. It reads production chat and voice conversations and pulls out behavioral failure signals: users rage-prompting (which is to say, swearing at your agent), rephrasing the same request five different ways, correcting the agent, asking for features that do not exist, or simply walking away after a response that was technically a success. Point it at your traffic and it generates the failure categories that actually apply to your product β broken workflows, retry loops, setup friction, churn risk β and then opens reviewed pull requests to fix them.
The framing is sharp because it names the gap the entire eval industry has been dancing around. Benchmarks measure whether the agent did the task. They cannot see the human on the other side deciding it is not worth the trouble. Task success and user success are different variables, and only one of them pays you.
The company came out of YC S26 with a 250 thousand dollar pre-seed backed by Entrepreneurs First, founded by Shubham Palriwala and Gourab Saha. The hard part is precision: frustration is a noisy signal, and a system that cries wolf on every rephrase becomes one more dashboard nobody opens. But the raw material is already sitting in every agent company's database, unread. Somebody was going to go mine it.
https://agnost.ai/
Comments