The first thing Max Brodeur-Urbas asked an autonomous AI agent to do was plan his girlfriend's trip to Italy. The machine returned a generic list of attractions, thinly researched and short on charm. He still thought the technology might change the world. Faith, after all, is often most vigorous just after the demonstration has gone badly.
It was late April 2023. AutoGPT, an open-source framework that prompted itself through multi-step goals, had become an online sensation. Its Discord server filled with people who wanted software to work on their behalf. Many of them could not get past the door. They did not know GitHub, dependency installation or the command line. Brodeur-Urbas saw the support requests and built a web interface that hosted AutoGPT in the cloud. React was new to him. He watched tutorials, made what he later called something usable but bad, and launched three days later.
The interface removed installation trouble. It also gave users a convenient new party to blame. When AutoGPT wandered, repeated searches or spent dollars without finishing, people complained to AgentHub, as the project was then called. Brodeur-Urbas investigated. The wrapper was not always broken. The autonomous agent beneath it simply could not do what the user had been promised by the surrounding excitement.
The useful disappointment
This was the more valuable prototype result. Users did not truly need an unrestricted digital creature improvising every move. They had recognizable sequences of work: research a company, parse a document, call an API, classify a message, place the answer in a spreadsheet. Most steps could be ordinary code. AI belonged at the narrow moments where language or judgment helped. Brodeur-Urbas and the small team began replacing autonomy with a pipeline people could see, arrange and correct.
“We needed to make AI automation reliable and affordable.”Max Brodeur-Urbas
There is an unfashionable quality to that insight. During a season of talking robots, Gumloop became useful by putting rails back on the railway. The visual workflow preserved the surprising reach of a language model while exposing the boring machinery around it. Boring machinery is underrated. It knows where Tuesday's invoice went.
Brodeur-Urbas had prepared for this kind of decomposition. He studied software engineering at McGill University, made the Dean's Honour List and helped lead the school's competitive programming group. Internships at Ericsson and Microsoft followed. In 2021 he joined Microsoft full time, working on CBL-Mariner, the Linux distribution used in Azure infrastructure. Competitive programming rewards the clean breakdown of a problem. Enterprise automation, it turns out, rewards much the same instinct, though the inputs are more likely to arrive by email and complain.
His student record contains a few clues about the range beneath that orderly engineering instinct. A Devpost project called Lo-Kate combined Android object recognition and voice interaction to help users navigate unfamiliar places, and won a hackathon prize. He also appeared in McGill's athletics record as part of a novice rowing eight. Neither project foretold Gumloop with the convenience of hindsight. Together they suggest a student willing to move between algorithms, interfaces and a boat whose success depended on eight people pulling in sequence.
A co-founder enters on cue
The company acquired its partnership with suspiciously tidy timing. On the morning of an AI meetup in Vancouver, an earlier collaborator quit and said AgentHub would not work. Brodeur-Urbas went to the event anyway. When organizers asked whether anyone else wanted to present, he raised his hand and improvised. In the recording, he says his co-founder had quit that morning and he is looking for another. The next voice belongs to Rahul Behal, a former AWS developer and fellow McGill alumnus, asking an enthusiastic question about the product.
Behal was contributing code within 24 hours. Brodeur-Urbas now jokes that he can barely watch the video because he sounds monotone. The footage is useful precisely because it has none of the polish later applied to founding legends. A nervous demo, a frank vacancy and one interested person in the room were enough. The coincidence opened the door; a shared appetite for building kept both men inside.
They started in a Vancouver apartment and worked on the same core problem from April 2023 onward. AgentHub entered Y Combinator's Winter 2024 batch. In May 2024 it became Gumloop, a name chosen with almost comic practicality. “AgentHub” was sometimes misheard when Brodeur-Urbas spoke quickly. The .dev address felt intended for developers. The product was no longer chiefly about autonomous agents. Gum suggested the sticky material connecting apps; loop described repeated automation; gumloop.com was available. Packs of gum could become merchandise. Corporate naming rarely receives such honest assistance from the confectionery aisle.
From bedroom build to enterprise platform
The company as a feedback loop
The operating habits were as iterative as the name. Early Gumloop tried to answer feature requests by the night's release, calling the practice same-day shipping. Every complaint went into a backlog. Confused clicks during demos counted as feedback. Small defects became “papercuts,” with Fridays reserved for clearing them. Brodeur-Urbas wrote that old demo videos should feel embarrassing within a month. Embarrassment, in this formulation, is not humiliation. It is a crude but sensitive instrument for detecting velocity.
Customer proximity did more than polish the interface. It showed where the product spread. One employee would automate a tedious piece of work, show colleagues, and become the resident AI person. Gumloop wanted that employee promoted. Three power users from Instacart, Webflow and Shopify eventually joined Gumloop itself, an unusually literal version of recruiting from the audience.
Brodeur-Urbas personally logged roughly 1,100 customer calls during the year after launch. He developed an internal diagnostic for the awkward moments: customer cringe. If he had to hedge, redirect or repeatedly tell someone where to click, the interface had failed to make its case. By 2026, that attention had been translated into a support operation where two people used Gumloop automations to run more than 500,000 support-related workflows each week. The company did not automate caring. It automated the sorting, enrichment and follow-through that gave two people more time to care.
The founders also used their own product to keep the organization small. They qualified signups, drafted outreach and analyzed support conversations through workflows. Gumloop announced its $17 million Series A in January 2025 as two founders, helped by interns and contractors but without other full-time employees. Brodeur-Urbas publicly described a ten-person, billion-dollar company as the aim. The arithmetic was memorable. Enterprise customers later made it less sacred.
Capital raised, by announced round
The Series A, led by Nexus Venture Partners, accompanied a move into San Francisco while the company retained its Vancouver roots. In March 2026, Benchmark led a $50 million Series B with participation from Nexus, First Round Capital, Y Combinator, BoxGroup, The Cannon Project and Shopify Ventures. Demand from larger companies required more engineers and a dedicated sales team. The ten-person cap yielded to the less poetic fact that businesses buying critical infrastructure expect humans to answer the telephone.
Agents return, with adult supervision
Gumloop eventually brought agents back. Models had improved, and the company had spent two years building integrations, execution infrastructure, authentication and controls. Its newer product lets employees create agents, connect them to workplace tools and deploy them through Slack, Teams or email. The visual workflow remains. So does the older lesson: capability becomes valuable when a person can inspect it, restrict it and understand what happened.
That has pushed Brodeur-Urbas into a more enterprise-shaped question. If every employee can build an agent, who sees the data it touches? Gumstack, the company's governance layer, records tool calls and data use across workplace agents, including some built outside Gumloop. The original project removed a technical barrier for strangers in Discord. The current one is trying to remove organizational barriers without also removing the locks.
“Understanding the task should be the only requirement for automating it.”Max Brodeur-Urbas
The aspiration is democratic in the old software sense: give more people the power previously reserved for specialists. Brodeur-Urbas wants the marketer, support operator or finance worker who knows a problem intimately to build the solution without first writing a specification for an engineering team. His public writing is equally clear about what stands in the way: learning curves, unreliable outputs, missing integrations, loose permissions and the deadening residue of software that technically works but feels dreadful.
He remains cheerfully suspicious of arguments that begin with a moat and end before anyone builds. Gumloop entered a market crowded with Zapier, n8n, established enterprise platforms and the model companies themselves. Brodeur-Urbas's answer has been to keep moving, listen closely and make complexity addressable one node at a time. This is not proof that the company will own the category. It is proof that practicality can be an ambitious philosophy.
The Italy itinerary was poor. The first interface was rough. The autonomous agents disappointed users. The company name had to change. The employee cap did not survive demand. At nearly every turn, progress required admitting that an attractive premise had met an inconvenient fact. Brodeur-Urbas kept the goal and revised the machinery. For a founder working with artificial intelligence, that may be the most human workflow of all.