Bernard Xie and his co-founder Brian Kim were brainstorming startup ideas during lunch breaks at their Asana office in San Francisco. They'd throw out concepts—opening a boba shop, building a golf app, planning events—but every time they tried to turn ideas into actual plans, they hit the same wall.
"We'd just end up typing stuff into ChatGPT," Bernard recalls. "There weren't any other tools that could help us break down our ideas into concrete projects and tasks."
Having both worked at Asana, they knew even the best project management tools couldn't bridge the gap between abstract ideas and actionable work. So they built their own demo over a weekend, applied to Y Combinator at 7:58 PM on deadline day, and somehow got in.
Kestral founders Brian Kim and Bernard Xie.
Three months later, they'd quit their jobs and start building Kestral.
Bridging the Sales-Product Gap
During YC, Bernard discovered companies didn't need help brainstorming—they needed help making sense of information they already had.
"Sales and R&D are very siloed," Bernard explains. "Maybe once a quarter the head of sales and the head of product will duke it out about what customers want versus what the roadmap is."
Kestral ingests sales transcripts, customer conversations, and internal docs to automatically generate contextualized tasks, creating an objective view of what work needs to be done—all without departmental politics.
Multi-Agent Architecture with Mastra
Initially, Bernard's team used simple OpenAI API calls where each step required manual confirmation prior to execution. As they ingested complex documents, they needed a workflow architecture that could manage itself.
"It didn't make sense for one agent to do it all.”
That's when they discovered Mastra. Sam's team was "grinding outside the YC office, handing out copies of Principles of Building AI Agents."
Using Mastra workflows, Kestral built a handful of specialized agents: one processes transcripts, another converts context into tasks, another organizes tasks into projects, and another prioritizes tasks based on overall company strategy.
Running on React, TypeScript, and PostgreSQL, the team hot-swaps between Anthropic, OpenAI, and Gemini models depending on tasks. When they hit issues with legacy workflows, the Mastra team provided real-time Slack support and introduced them to the new workflow beta, Mastra Workflows vNext.
Looking Forward
Bernard’s hope is that Kestral “becomes the default way to assign work, whether it's to agents or to humans—not just because we have the best quality tasks, but because we have the most context about how to accomplish them."
With three design partnerships from Series A to Series B companies, Kestral is building the infrastructure for how companies operate in an AI-native future.