Trishala Jain (@hitrishala) and Mark Nour launched Buildbox on August 3rd, pitching product teams an analytics layer for AI agents that appear healthy in evaluations and traces while still failing their users.
https://x.com/hitrishala/status/2084373120576815213 "Most AI agents don't fail where teams are looking," Jain wrote in an August 3rd thread on X. Her example is an agent that returns an answer and reports that it completed a task, even as the user repeats the request, tries to redirect the conversation or abandons it. (x.com)
The two-person San Francisco startup is part of Y Combinator's Summer 2026 batch. Buildbox says it analyzes user intent, agent behavior and outcomes, then groups breakdowns into ranked, reproducible issues. Its YC launch page says those issues can be tied to activation, conversion and retention, giving product teams a way to prioritize failures by business impact rather than by whether an error appeared in a trace. (ycombinator.com)
From agent traces to abandoned tasks
Most AI observability products begin with what happened inside a model-driven application. LangSmith, for example, lets developers inspect messages, tool calls, timing, token counts, errors and metadata across agent threads. Langfuse records prompts, model responses, tools, retrieval steps, latency and cost, alongside evaluation and session data. (docs.langchain.com)
Buildbox is placing its wedge further downstream. The relevant unit is the user's full attempt to complete a task, including moments when the agent technically responded but failed to produce the result the user wanted. Buildbox says it measures how often those patterns recur, identifies which users are affected and tests revised agent behavior or interaction patterns before presenting a proposed fix. (ycombinator.com)
That distinction matters because a clean execution record can still describe a poor product experience. An agent can call the right tools, avoid an exception and generate a plausible response while leaving the person on the other side unsure how to recover. Buildbox is betting that product managers will pay for a system that converts those ambiguous interactions into issues an engineering group can reproduce and address.
A broader UX automation bet
Buildbox's public website extends the pitch beyond conversation analysis. It describes vision-based agents that crawl a product, adopt specific user goals and move through flows such as onboarding, checkout and account setup. Buildbox then tests possible changes and can prepare a reviewed pull request, a ticket or a report for Slack. Buildbox says code changes remain subject to customer review and are not merged or deployed unless explicitly configured. (heybuildbox.com)
That broader product turns Buildbox into a proposed automated UX engineer, rather than a dashboard that only diagnoses an AI agent after deployment. The site says an initial audit can begin with a URL and test credentials, while deeper deployments can connect analytics, customer feedback, issue trackers and a scoped GitHub app. Access is currently sold through a demo-led process. (heybuildbox.com)
Jain and Nour met as UC Berkeley undergraduates and have built together for three years, according to their YC profile. Jain studied in Berkeley's Management, Entrepreneurship, & Technology program, combining electrical engineering and computer science with business, after product work at Google Labs, MongoDB and AppDynamics. Nour studied computer science and worked on distributed media-processing systems at Netflix, LLM-based customer-feedback analysis at Amazon and search-augmented reasoning at Berkeley AI Research. (ycombinator.com)
Buildbox is their second named product together. In 2025, Jain and Nour built Oratora, an AI interview coach selected for Mayfield AI Garage's Berkeley cohort. The program provided its selected groups with a stipend, mentorship and computing credits. Buildbox's terms of service identify Oratora, Inc. as the legal operator of the new product. (mayfield.com)
As a YC company, Buildbox falls under the accelerator's standard $500,000 investment: $125,000 for 7% and $375,000 through an uncapped safe with a most-favored-nation provision. The financing gives Jain and Nour room to test whether user-outcome analytics can become a distinct software category, or whether established observability and product analytics vendors will absorb the same workflow. (ycombinator.com)