Startup Patronus AI Inc. today debuted a tool called Percival that promises to help developers more quickly fix issues in artificial intelligence agents.

Patronus AI is backed by $20 million in funding from Datadog Inc., Lightspeed and other backers. Its flagship product is a platform that helps developers find the most suitable language model for an AI application, filter inaccurate output and perform related tasks. The company also offers evaluation datasets for testing AI applications’ reliability. 

AI agents often break down the tasks they perform into multiple substeps. There can be dozens or more substeps, which makes troubleshooting errors difficult. To determine why an agent performed a task incorrectly, developers have to identify the specific substep that caused the malfunction.

The workflow is further complicated by the fact that AI agent mistakes cascade. If a task’s fifth and sixth substeps rely on data generated during the third substep, an error in that data can cause them to malfunction. Such interdependencies make it more difficult to identify the root cause of errors.

Patronus AI’s new Percival tool uses AI to automate the process. According to the company, it can analyze the workflow through which an AI agent performs a task and identify the specific substep that is causing issues. Percival then generates a natural language summary that describes its findings.

Petronus AI says that the tool can troubleshoot more than 20 types of malfunctions. It can, for example, identify when an AI agent’s output doesn’t align with the user’s request or contains formatting issues. Percival also identifies situations where a prompt response contains out-of-date information.

Some tasks require AI agents to interact with third-party systems. Finding bugs in an application, for example, may require a programming agent to retrieve the application’s code from the GitHub repository where it’s stored. Percival detects errors that affect the third-party systems used for a task.

The tool spots when an agent uses the wrong external system to process prompts. It can also identify a range of related issues, such as cases where an agent picks the correct third-party application for a task but exceeds its usage caps. 

“When developers spend hours tracing through agent workflows only to find that a decision made five steps ago caused the final error, they’re not just losing time — they’re potentially losing control over their systems,” said co-founder and Chief Executive Anand Kannappan. “Percival gives developers the ability to instantly understand and fix their AI agents.”

Percival stores information about the AI agent errors it detects in what Patronus AI describes as an episodic memory. According to the company, the memory allows the tool to learn from past malfunctions and improve its detection accuracy. Additionally, developers can use the error data that Percival collects to benchmark how their AI agents’ reliability changes over time. 

Photo: Patronus AI