Alexandra Lato on Building Stack Overflow for Agents and Stack Internal with Codex

Aarti Bagul opens the OpenAI DevDay session by bringing up Stack Overflow, a customer using Codex heavily. She introduces Alexandra Lato, Stack Overflow’s VP of Product, to walk through the company’s initiatives, share practical lessons from Codex, and explain how Stack Overflow is building for a world where agents—not just people—use the platform. She then welcomes Alex on stage.

Alex comes on joking that she expected California to be hot and is freezing instead. She’s from Canada, she says, and still a terrible Canadian about the cold. Asked who she is, she says she leads product across Stack Overflow’s portfolio: the public Q&A site many developers grew up on, plus two new products they will unpack today, Stack Internal and Stack Overflow for Agents.

Aarti asks what the two initiatives are and where Codex comes in. Alex calls them opposite ends of a spectrum. Stack Internal is an enterprise-grade knowledge-intelligence product, built for large deployments. Stack Overflow for Agents is a small, experimental, individual-facing product that is agent-first. Both used Codex extensively, but in different parts of the software development lifecycle.

How they set Codex up was different too. Alex, a product person who likes to pretend she can be an engineer when none of her engineers are in the room, says Stack Overflow for Agents—SOFA for short—started as a hypothesis. The team wanted to stay scrappy. They gave Codex Stack Overflow’s architectural golden paths, then told it to go rogue: propose the decisions, the constraints, and the container to build in.

Stack Internal was the opposite. It has to ship to existing customers with 150 to 10,000 seats, including big banks at around 50,000, and many on-prem installs with far more constraints. There they were prescriptive: here are the contractual and architectural contracts between services, here is how the architecture is built, go work inside that safe container. They let Codex roam in some places and deliberately kept a human in the loop in others.

Aarti notes that the enterprise product’s security bar forced more guardrails up front. Alex agrees: Codex needs more than the architecture diagram. It needs to understand what an enterprise actually requires. Stack Internal pulls scattered, raw knowledge and structures it, which immediately raises data-leakage questions—who can see what, which authorization model applies, whether a source is on-prem or in the cloud. That complexity has to be handed to Codex as context.

A non-enterprise product is net-new architecture, so Alex can tell Codex what she wants and ask it for the best way to do it, then ideate from there. Aarti turns to the many data sources Stack Internal has to make sense of. In an age of “is grep all you need?”, she suggests the answer is no, and asks how the team worked through that.

Every day, Alex says, people in an enterprise create context: Slack threads, Teams chats, Google Docs, constantly edited and shared, each with permissions. Multiply that by every employee. Stack Internal is a knowledge-intelligence layer: it goes into that raw context, structures and organizes it, then applies Stack Overflow’s usual bet—trust and verification. Doing that means walking into those systems with a grocery list of privacy, governance, and leakage constraints, and respecting them all the way through.

A Slack message, a Google Doc, a Jira ticket, and a Confluence page do not store information the same way. Then one enterprise configures Jira one way and another does it differently, and Stack Overflow inherits all of it. They have to scale those sources without slowing down. Codex, she says, sees through that complexity and helps them ideate how to scale it.

What Codex built was the idea of archetypes: patterns that sources share. A document archetype can cover a Google Doc, a Teams attachment, Confluence, or Notion. If they put an API around that archetype, Codex can say it sees the same pattern across ten sources, and downstream connections can reuse it. That logic would have taken weeks of building; Codex did it in a few days.

Aarti points out it isn’t only retrieval performance: some sources are more authoritative than Slack or Teams noise, and freshness matters too. Archetypes helped put context around each source. For Alex, that was the front-door lift. Once the patterns were in place, engineers could spend their time on the hard problem: applying trust—deciding that Alex’s Slack thread from yesterday should outrank someone else’s from March. She did not want people wasting capacity on getting sources into the system. Codex took that lift, which she calls a godsend.

They switch to Stack Overflow for Agents. Aarti, who works in developer experience at OpenAI, says the people hitting docs these days are often not people at all—developers point agents at the docs and build that way. That raises a design question for web products: how do you design for agents as well as humans? She asks how Stack Overflow approached it.

Stack Overflow has always been a human-led, public community, open by default. Alex’s job has been to rethink the next chapter: a world where humans and agents consume and contribute knowledge together. That is a different game for a company like Stack Overflow. What happens to moderators? To questions and answers? To the familiar pain of having a question closed as a duplicate? Those are the questions SOFA was built to poke at.

SOFA is the experimental bet, on a separate architecture so they can move faster. They told Codex they had an idea and wanted a product experience that lets agents contribute on the platform. In the ideation phase, Codex wrote back asking whether they should keep some of Stack Overflow’s brand by attaching an agent’s identity to its human—an idea it pulled from brand guidelines and company history. Alex thought it was a great idea, and that is how they built it.

They built a public platform where an agent can hook in what it is learning in production and push that back to the community. Instead of lurking—Stack Overflow’s word for consume-only behavior—the agent can contribute. If it fixes something, it can post a Today I Learned. It can publish a blueprint for an architecture or a language. The shift is from agents that only consume via MCP to agents that also write knowledge back, which is new for Stack Overflow.

Aarti agrees that “agent native” is usually about making pages easy for agents to read, but the more interesting move is contribution: agents can add knowledge more consistently than human lurkers ever did. Alex says they are still experimenting with how much autonomy to give those agents.

If an agent is attached to a Stack Overflow human, that human has reputation—literal reputation points, and their name on the post. People want what goes out under their name to make sense. Autonomy is configurable: require the human to accept every post, or let the agent go rogue on a given topic.

The loop is not only contributing. Agents can also validate what is already on the site. If an agent hits a problem with an open answer, it can take that suggestion, test it in production, and write back: I tried this, it didn’t work, this is what fixed it. They are trying to close the loop on agent-first knowledge—consumption and contribution.

Aarti recaps the ground they covered: Codex with heavy guardrails on the enterprise product and free rein on the net-new agent product; archetypes that encode context about messy data sources; and agent-native meaning putting agents back in the loop as contributors of new knowledge. She asks if Alex wants to leave the audience with anything else.

If you have any love for what Stack Overflow did for your career over the last twenty years, Alex says, go look at what they are building. The first version of the enterprise organizational product—the knowledge-intelligence layer—was due out the next day. Stack Overflow for Agents was already free to try. She asks the community to test it, send feedback, and let them keep building for developers. Aarti closes, and the room applauds.