Companies across technology sectors are deploying AI agents to handle customer conversations, answer questions and complete tasks autonomously.
These systems promise efficiency at scale, but they also introduce a category of failure that traditional monitoring tools were not designed to catch. An agent can misunderstand a user, skip a workflow step or report completing a task it never performed without crashing, logging an error or raising a conventional alert.
“I build software that helps companies understand when their AI agents are failing in the real world,” said Pranav Bedi, founder and CEO of Moda. “Moda watches production conversations, identifies patterns like user frustration, incomplete workflows, or agents claiming they completed tasks they never actually did, and helps teams improve their agents based on real customer feedback.”
Bedi and his co-founder built Moda as part of Y Combinator’s Winter 2026 batch. The company has raised $6.1 million to build what it describes as a continual learning layer for production AI agents, giving teams visibility into how agents behave in real customer conversations.
Why production conversations reveal what tests miss
The gap between how an agent performs in a test environment and how it behaves in production can be wider for AI systems than for traditional software.
A prompt that works across a small set of curated test cases may fail unpredictably when exposed to thousands of real users with different phrasing, expectations and edge cases. Regressions can also appear without warning. A change that improves one type of response may silently weaken another.
The signals teams need, including user frustration, incomplete task resolution and hallucinated outputs, are not always captured by standard logging infrastructure. They are often embedded in the conversation itself.
“As more companies deploy AI agents into real workflows, software can now fail in less obvious ways,” Bedi said. “An agent may misunderstand a user, skip an important step, fail to complete a task, or create a frustrating experience without triggering a traditional error. Moda helps companies understand these failures in production and use real customer feedback to improve their software over time.”
Moda’s approach addresses an infrastructure blind spot the Y Combinator-backed company identified early. Rather than adding another dashboard of aggregate metrics, the platform analyzes agent logs to identify where each run broke, why it failed and whether the issue originated in the prompt, tools, workflow, model or product logic.
From raw traces to repeatable improvements
Once production conversations are captured, the next challenge is turning them into something actionable.
AI agents can generate enormous volumes of trace data. Every message, tool call and model response creates a record. Teams can quickly drown in information without a system that separates signal from noise.
Moda addresses this by automatically labeling failure patterns and surfacing them in a format engineering teams can use. The platform detects hallucination patterns, tool call errors, prompt injection attempts and signs of user frustration.
It also tracks regressions so teams can measure whether a change to a prompt or workflow improved agent behavior or made it worse. For teams running thousands of agent conversations each day, automated pattern detection can reduce the need to manually dig through individual traces.
In March 2026, Forbes named Moda one of the 21 most promising startups from Y Combinator’s latest batch, citing the company as part of a broader shift in which founders are building infrastructure for the AI-agent economy rather than directly for human users.
The road to reliable autonomous systems
The long-term trajectory for AI agents points toward greater autonomy.
Agents are expected to handle longer and more complex tasks with less human oversight. They will interact with other agents, make decisions based on ambiguous inputs and operate across multiple systems. As that happens, the cost of a single undetected failure rises.
“The big problem my work is trying to solve is how to make software better for everyone as AI becomes a larger part of how products are built, operated, and experienced,” Bedi said. “The people who benefit most are the end users relying on these systems, as well as the companies building AI products that need to be more reliable, useful, and trustworthy.”
Bedi’s background includes data infrastructure work at financial institutions, AI agent development at Shopify and early-stage startup operations at Cerebral Valley. He also won Cal Hacks 11.0 after building an interactive AI-powered storytelling application.
As more companies push AI agents into customer-facing roles, the teams that succeed may not simply be the ones with the most advanced models. They may be the ones that learn fastest from what their agents actually do in production.