.png){kind=logo}
A principal software architect whose fingerprints are on two of Microsoft’s open source AI frameworks, plus the connective tissue that keeps multi-provider AI usable.
Roger Barreto does not care about headlines. He cares about what still works years later.
“I do not have a shelf of trophies,” he says. “I have a trail of systems that are still running and frameworks that other people built careers on.”
This explains the through line of his work better than any title. Barreto is a principal software architect who has spent twenty-five years designing foundations, not features. He has built government platforms in Brazil that could not fail under peak load, created frameworks that entire companies adopted as their default, and then carried that same instinct into Microsoft’s current AI era.
He is originally from Brazil and is based in Dublin. He works in English and is a native Portuguese speaker. He has been a founder, a consultant, a team lead, and a mentor. His preference is simple.
“I care about the foundation underneath the feature,” he says. “The patterns. The abstractions. The thing the next ten developers will build on top of.”
Most software architects contribute to an open source project after it has a shape. Barreto helped define the shape.
At Microsoft, he was a core .NET engineer on Semantic Kernel from its inception, and later a founding architect on the Microsoft Agent Framework when it launched in 2025. Semantic Kernel is Microsoft’s orchestration SDK for AI. The Agent Framework is the next stage of how developers build AI agents. Barreto was present at the foundation of both, consecutively.
“Very few architects get to be there at the beginning of a major open source framework,” he says. “Being there twice means you do not get to hide behind what already exists. You have to design the reality.”
Inside Semantic Kernel, his work covered core abstractions such as streaming architecture, plugin systems, kernel hooks, and connector versioning. Those decisions affect how thousands of developers compose tools, manage multi-modality, and keep applications stable while upstream providers change behavior.
He also authored architectural decision records that governed why the framework works the way it does.
“I like documentation that makes the next decision easier,” he says. “If people cannot explain why a system behaves a certain way, the system becomes fragile.”
Barreto’s work gets most interesting where the industry gets messy. AI providers ship fast and often break each other’s assumptions. Each provider has its own model formats, tool calling patterns, and breaking changes.
Barreto built and maintained integrations across OpenAI, Google Gemini, Amazon Bedrock, Anthropic, Ollama, ONNX, HuggingFace, and Azure AI Inference. He also led the OpenAI SDK V2 migration across eight phases, a kind of effort that determines whether a platform keeps trust or loses it.
“Connector work is not glamorous,” he says. “It is maintenance under pressure. You wake up to a breaking change and you decide if the whole ecosystem stays coherent.”
He frames multi-provider stability as a responsibility to developers.
“A developer should not have to learn eight different mental models just to ship one product,” he says. “My job is to make the seams predictable.”
That same instinct carried into the Agent Framework work. Barreto created the Anthropic agent package that brought Anthropic models into Microsoft’s agent architecture through Azure Foundry. He also authored a suite of Foundry Agents tool samples, including Memory Search, Bing Custom Search, OpenAPI, MCP, SharePoint, Fabric, File Search, and Computer Use.
“Samples are not filler,” he says. “They become the pattern people copy when they are under deadline.”
Barreto’s path into AI did not come from a machine learning lab. It came from years of building for humans who needed software to work, even when the internet did not.
He wrote his first production code in a startup and later worked at Accenture in Rio de Janeiro while still an intern, building systems for oil and telecom companies. In Brazil, he moved from developer to architect to founder, repeatedly ending up in the same role, designing frameworks that other people could build on.
At Cast IT Group, he rebuilt a statewide education platform for Rio de Janeiro that handled test scoring and classroom systems under extreme concurrent load. He also built a high-performance phonetic search engine for Brazilian Portuguese, a problem he describes as the kind that offers no clean off-the-shelf solution.
At Go2web, he architected an offline-capable public school enrollment system for the state, capturing student data in schools with unreliable connectivity and syncing safely when connections returned.
“If the internet is a luxury, software has to be patient,” he says. “Data still has to be correct when the connection comes back.”
He also built an RPA framework at Wooza that became the company standard, and contributed early specification work on Microsoft Syntex eSignature, later shipped as a built in capability in Microsoft 365 for SharePoint and Word.
Barreto says the most durable work is not always a codebase. It is a team that can build well without you.
“At nearly every company I have worked for, mentorship and team formation were part of my role,” he says. “Not because it was required, but because a strong team outlasts your tenure.”
He describes standards, naming conventions, code review habits, and pairing as forms of structure. In his view, developer experience is not only API design. It is the culture that shapes how code gets written.
“You can build a great framework and still fail people if the culture is careless,” he says. “Quality is a daily practice.”
Barreto’s future goals are broad but consistent, career growth, work at the edge of knowledge, high impact at scale, and building products he believes in. He also signals a Silicon Valley chapter ahead, framed as the next environment where the rules of agentic AI tooling will be shaped.
He keeps the focus on the same idea that has guided him since he was twenty.
“I see my future at the forefront of AI technology,” he says. “I am not waiting for the next wave, I am redefining the parameters of what engineers and architects are doing before it happens.”
