A well-designed video game shows you how to play, rather than telling you.
Jesse Li and Isabelle Ilyia were friends from their undergraduate days at Georgia Tech. Part of YCombinator’s Winter 2025 batch with their startup Cedar, they explored various types of onboarding agents before realizing that they could make a generalized system to build onboarding agents.
Rather than generic tutorial videos or popover modals, Cedar's agents provide contextual, step-by-step guidance. In response to user questions, the agent can click through screens or enter in relevant info.
So instead of getting a canned workspace setup, they actually see that setup get created in real time.
Replacing Generic Onboarding with Intelligent Task Execution
Cedar's team needed to coordinate specialized agents for chat conversion, DOM manipulation, and workflow execution—all working together seamlessly at scale.
While their initial agent/workflow prototypes were in Python, when they started using Mastra, they had an immediate realization: "[The Mastra team] are super smart,” laughs Ilyia, the CTO. “They've thought of a lot of stuff that we [would] have to think of and implement ourselves.”
What they realized was that a wide swath of their challenges were less around their product and more around primitives — streaming, agent orchestration, and modular workflow construction.
"There are just certain things that everybody has to figure out, like streaming," Ilyia notes. "It's actually shocking to me that without Mastra, I would have to reinvent the wheel on most of these things when it's like, this is a solved problem."
The migration allowed Cedar to think "one level higher that's actually tied to the product we're developing rather than agentic development at the lowest level."
A Multi-Agent System Orchestrated by Mastra
Cedar has implemented what they describe as an "agent army": a sophisticated multi-agent system where each agent specializes in specific interactions, all coordinated through Mastra workflows acting as an orchestration layer for agent handoffs.
When a user asks a question, Cedar's main chat agent responds, then spawns a cursor with step-by-step UI guidance. This involves multiple specialized sub-agents: one that writes UI-based instructions, another that understands the current DOM state, and a third that pinpoints specific elements for user interaction.
"We have an overall supervisor that orchestrates the army," Ilyia explains. "It's like, 'hey, you go get me this,' and then once I have this information, now I need [another agent], who's an expert in the DOM, to break down each instruction and tell the user what to do."
Infrastructure Problem Solved, Product Focus Achieved
The impact on Cedar's development velocity has been transformative. Without Mastra, Ilyia thinks they would "probably still be in Python with duct tape and bubblegum around pieces of how people solved different problems."
Cedar's customers now benefit from multiple coordinated interactions, chatbots, cursors, and proactive suggestions that work together seamlessly. "Having all those interactions orchestrated in a way that is really powerful wouldn't be possible without the modular pieces that Mastra gives us," Ilyia explains.
The "batteries included" philosophy allows Cedar to focus on their core product rather than infrastructure. "I don't know if there are any alternatives that do what Mastra does," Ilyia notes. "There is no other product that orchestrates modular concepts in the [same] way."
Looking ahead, Cedar is expanding their open-source approach to enable even deeper software integration, building tools that will allow AI copilots to make actual changes in applications.
With Mastra handling the complex infrastructure, Cedar can concentrate on their own customer onboarding.
"Now we can just focus on the business side," Li reflects. "Just partnering and getting [customers set up]."