Voker, a YC Summer 2024 company founded by Alex and Tyler, launched an agent analytics platform that reveals critical insights about production AI systems. According to their survey of YC founders, 89% of production AI agent systems handle fewer than 10,000 conversations per month, and over 90% of teams only discover agent failures through customer complaints.
Survey Reveals Visibility Gap in Production Agent Systems
Voker's research among YC founders identified a significant operational problem: the lack of proper visibility into agent performance in production. More than 90% of respondents reported that customer complaints are their only signal for detecting agent failures. Even teams already using existing tools like LangSmith, Langfuse, or Grafana reported that understanding and improving agent behavior remains their hardest operational challenge.
The company's thesis positions 2025 as the year agents reached production deployment, with 2026 being the year teams must focus on hardening and optimization. Teams currently push prompt changes hoping to fix problems without breaking existing functionality, creating a reactive cycle of fixes without systematic understanding of impact.
Voker Provides Dynamic Categorization Without Manual Configuration
The platform works through a lightweight SDK that is LLM stack agnostic and purpose-built for agent products. Voker processes LLM calls by automatically annotating individual conversations, identifying user intent and corrections without requiring manual setup. The system uses LLMs and hierarchical text classification to create dynamic categories, providing higher-level insights without requiring users to read individual conversations.
Crucially, Voker does not use LLMs for core data engineering tasks like processing events or calculating statistics. This architectural decision ensures analytics remain consistent, reproducible, and accurate—addressing a key limitation of simply using Claude or ChatGPT to analyze logs, which the team argues produces inconsistent statistics and overfits some insights while underfitting others.
Platform Tracks Three Core Analytics Primitives
Voker's analytics focus on three fundamental metrics: Intents (what users want from agents), Corrections (when users must correct the agent), and Resolutions (whether intents are eventually resolved). This framework allows agent engineers, AI product teams, PMs, analysts, and business teams to access digestible insights without creating bottlenecks or delays.
The platform offers SDK support for Python and TypeScript, wrapping LLM calls to OpenAI, Anthropic, and Gemini. Pricing starts with a free tier of 2,000 events per month requiring email signup, with paid plans beginning at $80 per month and including a 30-day free trial.
Voker launched on Hacker News on May 12, 2026, receiving 55 points and 22 comments. The company provides demo videos and documentation at app.voker.ai/docs.
Key Takeaways
- Voker survey found 89% of production AI agent systems handle fewer than 10,000 conversations per month
- Over 90% of YC founders report discovering agent failures only through customer complaints, indicating a major visibility gap
- The platform uses a lightweight, LLM-agnostic SDK that automatically annotates conversations and identifies user intent and corrections
- Voker deliberately avoids using LLMs for core data engineering to ensure consistent and accurate analytics
- Pricing starts with a free tier of 2,000 events monthly, with paid plans at $80/month including Python and TypeScript SDK support