Both labs are showing up.
Anthropic and OpenAI are joining the AI stage at TechCrunch Disrupt 2026. That’s the announcement, and it’s a short one. But if you build bots for a living, as I do, a shared stage between the two labs whose APIs probably sit in your .env file right now is worth thinking about carefully.
I want to be honest about what we know versus what we’re guessing. We know both companies are confirmed for the AI stage. We know Amjad Masad, CEO and co-founder of Replit, is on the Disrupt Stage. We know applications are open to host Side Events, and that Audience Choice voting is live to decide which sessions advance to the TechCrunch Founder Summit. Everything past that is inference. I’d rather give you a useful listening guide than a fake preview.
Why a shared stage matters more than a keynote
Solo keynotes are marketing. Shared stages are different. When two labs present into the same room, the same audience, and the same press cycle, they end up implicitly answering each other. Developers get a rare side-by-side read on how each company frames the same problems: agent reliability, tool calling, long-context behavior, cost per task.
That framing is the part I care about. My architecture decisions rarely hinge on benchmark deltas. They hinge on whether a provider treats agents as a product surface or as a capability you assemble yourself. Those are two genuinely different philosophies, and they lead to two genuinely different codebases on my end.
The questions I’d want answered
If I had a mic in that room, this is what I’d ask, and it’s what I’ll be scanning transcripts for afterward:
- Agent failure modes. Not the happy path. What happens on step fourteen of a twenty-step task when a tool returns garbage? Retry semantics and error surfaces are where agent projects die.
- Deprecation policy. Model retirements break production bots. Any concrete commitment on notice periods and pinned versions is more valuable to me than a new capability.
- Cost predictability for long-running agents. A bot that runs for hours has a fundamentally different billing shape than a chat turn. I want to hear both labs address that directly.
- Where the abstraction line sits. Are we writing orchestration ourselves, or are we adopting theirs? This determines portability, and portability determines negotiating power.
- Evaluation. Everyone ships agents. Almost nobody ships a credible story about how they know the agents work.
The Replit signal
Masad’s presence on the Disrupt Stage sits alongside the lab announcements in a way I find more interesting than either alone. Replit occupies the layer between model providers and shipped software. That layer is where most of the actual friction lives for builders. Model quality has been improving faster than our tooling for deploying, monitoring, and rolling back what models produce.
So my hope is a stage that isn’t only about capability claims. The gap between a working demo and a bot you’d point a paying customer at is mostly plumbing, and plumbing is where builder-facing companies speak more plainly than labs do.
Agents are getting funded on operating, not building
One more data point from the same news cycle. Runable raised $21M on a bet that AI agents can go from building businesses to growing them. Read that framing twice. Building is a bounded task with a visible end state. Growing is open-ended, continuous, and requires an agent to make judgment calls without a human confirming each one.
That’s a much harder engineering problem, and it maps directly onto the questions above. Continuous operation means state management, memory that doesn’t rot, and guardrails that hold up over days rather than minutes. When capital moves toward always-on agents, the demands on model providers shift too. I’d like to hear both labs speak to durability rather than peak capability.
What to actually do about it
Practical suggestions, since this is a build-focused site. Vote in the Audience Choice poll if there’s a session you want at the Founder Summit; these things are decided by small numbers of engaged people. If you run a local builder community, the Side Event application is open, and side events are consistently where the useful conversations happen.
And in the meantime, write your integrations behind an interface. Two labs on one stage is a healthy sign for developers, because competition on the same day in the same room tends to produce clearer answers than either would give alone. Your job is to stay in a position where you can act on those answers without a rewrite.
I’ll be reading the transcripts closely. Not for the announcements, but for the parts where someone gets asked a hard question about reliability and has to answer it in front of the other lab.
🕒 Published: