OpenAI DevDay 2026: AI Is Starting to Feel Like a Teammate

I watched OpenAI DevDay 2026 expecting new models, better performance, and more developer tools.

We got all of that. But what stayed with me wasn't any single announcement.

It was how everything is starting to connect.

Models, AI agents, computers, workspaces, developer tools, and distribution.

A while ago, I wrote about Grok Bot and how AI agents were starting to feel like teammates. After watching DevDay, that idea feels a little closer to reality.

Dots and ChatGPT Space: AI Inside the Workflow

One of the most interesting announcements was Dots.

Unlike a traditional chatbot that waits for your next prompt, a Dot can take a goal and continue working using its own cloud computer, browser, and connected tools.

OpenAI also introduced Specialist Dots for tasks like accounting, email marketing, legal analysis, and ticket resolution.

That made me wonder: instead of one AI assistant doing everything, could we eventually work with a team of specialized agents?

Then there's ChatGPT Space, where people, agents, and shared project information can work together.

Imagine discussing a feature requirement and simply saying:

“@dot, check which parts of the project this change affects.”

Instead of copying information between ChatGPT, documents, and project tools, the agent could investigate directly within the workflow.

That's the shift I find interesting.

AI isn't just answering questions about our work. It's starting to participate in the work itself.

The Research Numbers Caught My Attention

One of the most interesting slides wasn't about a product.

OpenAI showed how its researchers were increasingly using agents for longer tasks.

For tasks estimated to take humans 8–16 hours, the reported success rate with zero human intervention increased from 10% in January to 35% in July 2026.

35% is still far from reliable enough to blindly trust.

But the metric itself is interesting.

We've spent years asking:

“Can AI give the right answer?”

Now we're starting to ask:

“How long can AI keep working before a human needs to step in?”

That's a different way to think about AI capability.

The Developer Platform Is Coming Together

DevDay wasn't only about Dots.

OpenAI also introduced GPT-6.1 Sol, focusing on balancing performance and cost, alongside Ultrafast, which was presented as reaching up to 300 tokens per second.

Those improvements matter when an agent needs to make repeated model calls, use tools, test results, and keep working.

On the developer side, Codex Harness and the Agents API make it easier to build agents with execution environments, context management, tools, and longer-running workflows.

There were also demonstrations of agents navigating applications, writing code, working from sketches, and interacting with hardware.

Not every live demo went perfectly, which was a useful reminder that these systems are still evolving.

And with plugins, third-party integrations, and an OpenAI marketplace, there's also a distribution layer taking shape.

OpenAI summarized its platform direction in three parts:

Models → Building → Distribution

To me, that was probably the clearest explanation of the entire event.

What This Means for Software Engineers

The model is becoming just one part of a much bigger system.

A capable AI agent also needs context, tools, permissions, an execution environment, monitoring, and human approval.

Giving an agent access to documentation is one thing.

Letting it modify production systems, send emails, or make financial decisions is something else.

So as agents become more capable, I think the engineering challenge isn't simply making them more autonomous.

It's figuring out how much autonomy we should give them, and where humans should remain involved.

My Takeaway

I don't think the biggest story from DevDay was GPT-6.1 Sol, Dots, or Codex individually.

It was how they're starting to fit together.

Models provide intelligence. Agents use tools and computers to act. Workspaces let humans collaborate with them. Developer APIs and integrations make it possible to build and distribute new experiences.

We're moving from asking AI for answers toward giving AI work to do.

And that leaves me with one question:

What happens when software isn't just something humans use, but something AI can use on our behalf?

That's what I'm most interested in exploring next.

Happy coding. ☕️