When longtime Retool builder Matei Canavra started his bootstrapped recruiting business, he knew he eventually wanted to create a feature that automated candidates’ engagement with the companies they were applying to. His business automates job searches, helping tech talent get hired faster through a combination of AI-powered automation and high-level career strategy. But finding and applying to so many roles can be tedious and trying for candidates, so he hung onto this idea of a “Super Recruiter” for over a year. Then came Retool Agents.
In our latest Builder Talk, Kent Walters—lead PM for Agents at Retool—sat down with Matei to explore how he’s deployed Retool in unexpected ways, from Chrome extensions to Electron apps, and how Retool Agents finally made his most ambitious automation ideas possible.
Matei didn’t wake up one morning and decide to put Retool apps inside Chrome extensions when he was first building his recruiting platform. The journey was much more organic, and especially instructive for anyone building with AI today.
“We started just out of Notion,” Matei explains. “I was going through my own job search back in 2022… you got to get a lot of applications out there. In the spirit of not building things you don’t need, I just used Notion because it had the UI.”
That Notion setup got shared around to other job seekers. People liked it, but it wasn’t scalable. Matei tried to make it work with Notion templates, duplicating setups for each customer. Eventually, it became untenable, and he jumped to Retool.
It’s not about the specific tools, but the progressive validation. Matei started with the simplest thing that could work, then systematically addressed limitations as they became real bottlenecks. The real lesson: don’t architect for problems you don’t have yet.
“[Our product] helps tech talent find a job faster by automating the job search and offering coaching, branding, strategy,” Matei says. “You can imagine that there is a lot of human in the loop, a lot of automation, and a lot of customer service. It’s pretty hard to find a platform that handles all of it, but with Retool we’re able to piece it together”
Today, Matei’s business runs on Retool in a number of ways:
- Candidate portal: A polished external app that has the look and feel of a custom React app.
- Internal operations tools: Embedded Retool apps inside Electron for sending out applications.
- Chrome extension: Retool iframes for job collection on LinkedIn.
- Workflows: Complex data pipelines for job curation and AI review.
- Retool Agents: The missing piece for non-deterministic automation.
For over a year, Matei had teased the “Super Recruiter” feature—an AI agent that would handle job prioritization on behalf of candidates. But until Agents arrived, it was technically impossible to build reliably with the tools he had available.
“This is something that we haven’t been able to build for about a year because it'd be just too complex to do,” Matei explains. “And now Retool Agents enable us to do that very, very quickly.”
The business problem was real: candidates needed to wake up early (sometimes 5 AM) to manually prioritize their job queue before the operations team in the Philippines would start sending applications. For a service that was supposed to save time for already stressed out candidates, the experience wasn’t great.
With Agents, Matei built a three-tier agent system:
- Super Recruiter (parent agent): Takes high-level requests via voice or text.
- Scout agent: Searches for new jobs based on resume and objectives.
- Prioritizer agent: Interfaces directly with the job queue and reorders based on candidate preferences.
Rather than building one monolithic agent, Matei created specialized agents that can be composed together. The Super Recruiter acts as an orchestration layer—almost like a Model Context Protocol (MCP) server with natural language input—while the specialized agents handle specific domains.
When the conversation turned to MCP, Matei shared a perspective that goes beyond the typical “wrap your API endpoints” approach most developers are taking.
“Something that is agentic... knows when to terminate its own set of actions. I’d add to that that it should also know where to begin,” Matei explains. "My ideal version of an MCP server is like, ‘hey I need you to go solve this problem’—I’m passing problems and I’m getting solutions.”
Rather than exposing individual endpoints like “get teams,” “create issues,” “batch update,” Matei advocates for higher-level operations like “start a new cycle” or “send cycle update.” These operations might do multiple things in the background, but they match how teams actually work.
For his MCP server, instead of just basic CRUD operations, he built workflow-level actions:
- “Start new cycle” (creates project, sets up milestones)
- “Send cycle update” (handles project updates, milestone updates, various notifications)
- “Kickoff project” (orchestrates multiple backend operations)
This approach reduces the cognitive load on agents—they work with business concepts rather than implementation details.
One of the most practical questions that came up during Kent and Matei’s conversation was when to build a Retool Workflow versus when to build a Retool Agent. Matei’s answer was refreshingly pragmatic:
“If you want to be lazy, you probably want an agent. And if you don’t care as much about the accuracy, and you care more about getting something done and iterating with it... if it’s faster to actually have a conversation to iterate, probably reach for an agent.”
More specifically:
- Use Workflows when: You know the starting point, the steps, and the end point. The process is deterministic and you want maximum reliability.
- Use Agents when: You want to iterate conversationally, or when the specific process should vary based on context and judgment.
This framework isn’t either/or, and that’s what makes it powerful. In Matei’s system, the agents call workflows as tools. The agent provides the intelligence about which tool to call and when, while the workflows provide reliable execution of known processes.
“Those tools that I provided to my agent—those were workflows. The agent selected which tool to call, and that tool itself is the abstraction over those predefined steps,” Matei says.
Perhaps the most striking thing about Matei’s approach is how thoroughly it’s driven by actual business needs rather than technical novelty. Every piece of his architecture—from the backend-for-frontend pattern to the agent hierarchy—serves a specific operational requirement.
The Chrome extension exists because they need to collect jobs from LinkedIn’s existing infrastructure. The Electron app exists because they need desktop automation with human oversight. The agent system exists because manual job prioritization was creating real friction for customers.
“A big part of the value proposition here is that we are solving for time,” Matei explains. “So to ask people to do more in our own system is kind of antithetical to what we're providing as value prop.”
This is the difference between building AI for AI’s sake and building AI to solve real problems. Every automation decision gets filtered through: does this actually save time and reduce friction for our customers?
As Kent and Matei wrapped up their conversation, a clear throughline emerged in Matei’s approach offering actionable insights for anyone looking to build smarter, more effective internal tools powered by AI.
- Start with manual processes that already work. Don’t try to automate broken workflows—fix them first, then automate the good version.
- Build incrementally. Matei didn’t start with agents. He started with Notion, moved to Retool apps, added workflows, built APIs, and finally added agents when they solved a specific problem he couldn’t address otherwise.
- Focus on composition. The most powerful systems aren’t monolithic agents—they’re specialized agents calling purpose-built workflows, with proper abstractions at each level.
- Optimize for the real bottlenecks. The most impressive technical solution isn’t always the most valuable business solution. Sometimes the biggest impact comes from eliminating a manual process that requires your users to get up at 5 AM.
- Keep humans in the loop where it matters. Matei’s system automates job discovery and prioritization, but keeps humans involved in actually sending applications. Know where precision matters and where “good enough” is actually good enough.
Whether you’re building internal tools for your own company or creating products for external customers, Matei’s approach offers a clear path: understand the real work that needs doing, build systems that reduce genuine friction, and use AI as a tool for orchestration rather than an end in itself.
Want to see Matei's full demo and hear more about his process? Watch the full episode here.
LinkedIn: Matei Canavra
Explore Matei’s project built with Retool Agents: https://magicapply.jobs/
Reader
