Startup Patronus AI Inc. has unveiled Percival, a tool designed to help developers address issues in AI agents more efficiently.
The company secured $20 million in funding from backers like Datadog Inc. and Lightspeed. Its primary offering is a platform that assists developers in selecting the optimal language model for AI applications, filtering inaccurate outputs, and handling related tasks. Additionally, the company provides datasets for testing the reliability of AI applications.
AI agents typically break down their tasks into multiple sub-steps, sometimes numbering in the dozens. This complexity makes it challenging to identify errors. To understand why an agent failed, developers need to pinpoint the specific sub-step responsible for the failure.
Error propagation complicates workflows further. For instance, if steps five and six rely on data generated in step three, an error there could cause subsequent failures. Such interdependencies make root cause analysis even more difficult.
Percival automates this process using AI. According to Patronus AI, the tool analyzes workflows of AI agents and identifies problematic sub-steps. It then generates a natural language summary of its findings.
Patronus AI claims that Percival can detect over 20 types of issues. These include mismatches between AI outputs and user requests, formatting problems, and instances where responses contain outdated information.
Some tasks require AI agents to interact with third-party systems. For example, debugging an application might involve retrieving code from a GitHub repository. Percival can identify errors impacting these third-party integrations.
The tool can detect when an agent uses the wrong external system or encounters related issues, such as exceeding usage limits despite choosing the correct application.
"When developers spend hours tracing workflows only to find that a decision made five steps earlier caused the issue, they're not just wasting time—they may be losing control of the system," said Anand Kannappan, co-founder and CEO. "Percival empowers developers to quickly understand and fix their AI agents."
Percival stores detected errors in a system called episodic memory. This feature enables the tool to learn from past failures and improve accuracy. Moreover, developers can use the collected error data to benchmark the reliability of their AI agents over time.
