The first version of Arcade was supposed to watch a website break and figure out why. An alert would arrive; an AI agent would inspect servers, storage, observability tools and code; then it would diagnose the failure and perhaps repair it. Alex Salazar and Sam Partee believed they were building a site-reliability agent. They got it to work. Prospective customers still did not care enough.
The awkward part was that engineers kept interrupting the demo. Their questions were not about the agent on the screen. They wanted to know how it had authenticated into all those separate systems. How did it call the tools? Where did the credentials go? Could it act for a particular user without becoming an overpowered service account?
The supporting machinery was getting more attention than the product it supported. Salazar and Partee eventually threw away the visible idea and kept the plumbing. That plumbing became Arcade.dev, a San Francisco company working on the action layer between AI models and the services where people actually do work.
The hallway outside his bedroom was an office
Salazar grew up close enough to the computer business to hear it through the wall. His mother built companies that distributed IBM, HP, Compaq and other computer products into Latin America and the Caribbean. Her second company began in the family garage. Calls and faxes passed through the hallway beside his bedroom. He remembers knowing about server trouble, firings and currency devaluations before he knew how to drive.
There was labor, too. Salazar has recalled stuffing envelopes for mailings at 10 and cold-calling prospective computer buyers in Africa at 15. The family business gave him unusual access to parts catalogs. He assembled his own gaming rigs, ordering memory, processors and graphics cards directly. The fascination was not yet a career plan. He liked computers because he liked games.
For a while, he expected to become a pediatrician. Then advanced biology collided with a mind that disliked memorization and exceptions. Physics felt different. During college, he once arrived unprepared for an electromagnetics exam and reasoned his way to a B. The experience helped point him toward engineering, then computer engineering, then software. After starting at Florida International University, he transferred to Georgia Tech to study computer science.
“I’m still trying to live up to my mom’s shadow.”Alex Salazar, recalling his entrepreneurial upbringing
The dot-com crash turned graduation into another forced experiment. Software jobs had contracted. Salazar says he crashed roughly 40 campus interviews that had not invited him, hoping persistence would create a slot. Microsoft and IBM both eventually called with offers in sales, which initially offended the engineer. IBM showed him the compensation. He needed rent money. He accepted.
He stayed for four and a half years and loved the work. Later, he would call the experience more valuable than his Stanford MBA. Sales taught him how a technical product travels through an organization - how a buyer describes pain, how a budget appears and why good engineering does not automatically produce a purchase. That bridge between product and market became the recurring advantage in his career.
Software engineering
Customer language
Okta
Arcade
Identity before agents had identities
While at Stanford Graduate School of Business, Salazar wrote a paper about a coming shift in cloud computing. He argued that a large, well-capitalized company would offer computing and storage as services, setting off demand for a new generation of databases, security products and middleware. When Amazon’s EC2 and S3 made the prediction tangible, he decided to build in that new layer.
The result was Stormpath, which Salazar co-founded in 2011 and led as CEO. It offered authentication as an API for developers, treating identity as infrastructure that product teams could use instead of rebuilding. Okta acquired Stormpath in 2017. Salazar moved into Okta, where he led developer products and product strategy, then became general manager for new-product introductions.
Arcade’s account of his Okta years attaches numbers to the work: developer products representing 25 percent of total bookings, a network of more than 5,000 authentication integrations, and a new authentication-centered proxy product that reached $9 million in revenue during its first year. The deeper continuity is conceptual. Stormpath asked how an application knows who a person is. Arcade asks what should happen when an application can act for that person.
The investor who still felt like a founder
After Okta, Salazar did not rush into another company. He had seen founders force ideas simply because they wanted to be founders. His view was blunt: choosing the opportunity is most of the game. He discussed a 24-month window with his wife and planned to immerse himself in engineering teams and customer problems - to put himself, in his phrase, “in the path of luck.”
The pandemic altered the plan. He joined Neotribe Ventures in 2020 and intended to make venture capital his permanent work. To improve his deal flow, he followed advice from Benchmark co-founder Andy Rachleff: meet strong technical people in the category, ask each whom they respect, then keep following the chain. In AI, that network eventually led to Partee, a Redis engineer who became Salazar’s technical sounding board for investments.
Then GPT-3.5 and ChatGPT arrived. Salazar recognized the shape of a platform shift. Models would sit beneath a fresh layer of applications; those applications would need middleware. He and Partee saw teams building too much of that connective tissue themselves. Salazar left Neotribe, which became Arcade’s first investor, and called Partee with the idea of starting together.
“I realised I was really still a founder and not a VC.”Alex Salazar on leaving venture capitalThe pivot
Two people, no product code, several wrong answers
Experience made Salazar more willing to begin small. The founders rented a tiny room in a coworking office. He did not take a salary. They accepted only enough money to pay Partee and refused venture meetings while they searched. Their rule was severe: no product code. Instead, they mocked up products, put them in front of potential customers twice a day and asked whether anyone would pay $100,000.
Stealth, in this version, was not a curtain around a grand reveal. It removed the cost of changing direction. The pair could be one company on Monday and another on Thursday without explaining the old pitch to investors or the market. The site-reliability agent failed. Other concepts failed after it. Salazar has even described testing one deliberately bad idea in the hope that the reaction would produce an insight.
The turning point came when authentication and tool calling started pulling customers forward. An AI model could draft an email but could not safely send it as the right person. It could suggest a CRM update but needed permission, credentials, an audit trail and a reliable tool to make the change. These were familiar identity problems in an unfamiliar architecture.
Two founders could discard an idea without dragging a large organization behind them.
Asking about a $100,000 purchase made polite enthusiasm less useful.
The questions that stopped the demo exposed the product customers needed.
The first agent could fail while its underlying infrastructure survived.
A smart intern still needs a badge
Salazar often explains agents through the gap between intelligence and authority. Language models are probabilistic. They can choose the wrong tool, invent a parameter or misunderstand an instruction. An enterprise permission check cannot share that temperament. The model may reason creatively; the system deciding whether it can send, delete, transfer or update needs enforceable rules.
Arcade therefore sits between agents and business systems. Its runtime handles credentials away from the model, connects an action to the user on whose behalf it is happening, applies policy and records what occurred. Salazar frames autonomy as a gradient. Letting an agent categorize an inbox carries a different risk than letting it initiate a wire transfer. Human approval can remain in the loop where value and risk demand it.
Arcade announced a $12 million seed round in March 2025. In June 2026 it announced a $60 million Series A led by SYN Ventures, with strategic investment from Morgan Stanley and Wipro, bringing total financing to $72 million. The company says it now ships more than 8,000 agent-optimized tools, and that its team authored the MCP authorization specification adopted by Anthropic and major MCP clients and servers.
Funding makes the ambition larger, but the central idea remains ordinary enough to picture. Work happens across Gmail, Slack, GitHub, Salesforce, databases and internal systems. An agent becomes useful when it can cross those boundaries. It becomes deployable when each crossing has a known identity, a narrow permission and a record.
Salazar’s career has circled that boundary for more than a decade. The child with the parts catalog learned how computers fit together. The reluctant seller learned how companies decide. Stormpath handled the login. Okta showed identity at enterprise scale. Arcade is the next question in sequence: once software can speak for us, who lets it use its hands?