Every week, there is a new startup building its own harness. But the best harnesses will be built by the labs. They control the models and can optimize both together. They also have a strong incentive to build the best harness. So if you’re building a harness, we’re now back to startup 101: pick a niche. Not too small so there is enough revenue. Not too big so you don’t get outcompeted by a giant with more resources and better distribution.

@OpenAI just launched the Data agent in ChatGPT Work, with @DeepnoteHQ available as a native connection. Connect Deepnote to bring the data sources and context your team already uses into the conversation, then explore your company’s data, build interactive data apps, or take action. To get started, add the Data Plugin and Deepnote in ChatGPT Work, and start your exploration: vist.ly/5iin8 Special thanks to Charles Wang, Steve Imm, Oliver C., and Shaina Li for the collaboration.

In a year, you won't know which model you're running. Fifteen years ago, people looked for a Pentium or Celeron sticker before buying a laptop. Then chips got fast enough that the sticker didn't matter anymore. Models are heading the same way. Most tasks don't need the frontier, and routing to a cheaper model saves a lot of money. As a result, I believe that by the end of the year we will no longer pay per token but per some obscure intelligence unit. ChatGPT is currently testing hiding models from you, with Sol as a default and lower-capability models being routed to. I think this will soon become the new norm for all providers. Do we need one more black box in AI?

OK guys, be honest, who's distilling models en masse again?

It’s getting harder to tell a larger model from a smaller one. Once the output is comparable, speed becomes the new differentiator. The intelligence gap is closing, and at some point, a smarter model isn’t worth waiting for. Everyone says that Google is too far behind to catch up, but I think they are building for this.

I was on a plane with bad wifi and 15% battery left, so I told my agent to hurry up. Best experience I've had with it in months. It makes sense. Benchmarks are optimizing for the highest score, not my time. So every model is incentivized to lint, typecheck, spin up a browser, take screenshots, typecheck again. Great for fixing bugs and maxxing out benchmarks, not so great for building. Video credit: @blended_jpeg

POV: my four agents merging into main at the same time

Should agents decide their own workflow? Besides loops, maybe it’s one of the biggest debates in AI engineering right now, and we're having it inside Deepnote too. Non-deterministic won’t fly in enterprise, which is where we see a lot of adoption: some workflows need guaranteed, auditable execution patterns. Good thing is, we’ve architected Deepnote to enable both types of workflows. Notebooks are actually pretty good at that → isolated environments where non-deterministic runs can happen (and plain English instructions can be beautifully formatted) + deterministic behavior via code, with a clear sequential/reactive pattern.

Should we stop building UI entirely? A legit question I catch myself asking as I build Deepnote. We attracted strong demand initially due to having a much better UI vs incumbents. I'm increasingly unsure if any software company should keep shipping more screens. Work already starts in Claude Code, Codex, or Cursor. For better or for worse, the chat is the front door now. But whatever gets built there still needs a place to connect to data, deploy, run, collaborate, and govern access. Giants like Salesforce are embracing being headless for agents. If you think about it, each SaaS product eventually boils down to being an API or a database. Maybe we should take this into account as we build out the companies.

AI vendor lock-in is dying before it ever shipped. Every company I talk to is building the same thing: a CRM, an internal data analyst, a productivity tracker. Think of this as an operating system custom-made for your business — on top of whatever model is cheapest that week. Whoever hosts this operating system will have the moat. Right now, that's increasingly the companies themselves. For OpenAI and Anthropic, this is both a threat and an opportunity. If companies build this first, they can switch to another model anytime, including the open source models. If OpenAI/Anthropic build this first, they’ll finally get some vendor lock-in.

Agent Harness > Model The model leaderboard reshuffles every week. Teams that hardwire themselves to one lab spend every release cycle re-evaluating their whole stack (and life choices). Teams that own their harness can just flip the switch. We felt this at Deepnote. We once moved back a model generation because A/B tests with real users preferred the old one. The swap took minutes. Everything that made the agent useful stayed: the context layer, permissions, traces. That's the harness. And it's where the compounding happens. It knows your schemas, your permissions, your definition of done. If configured correctly, it can even outperform the vanilla harnesses by the frontier labs on ARC-AGI-3 (as shown by the @PrimeIntellect team). The labs know this. It's why every one of them now ships compound systems, and not just API-driven models.

At @DeepnoteHQ, we're very data-driven. Also, the data in question:

🚨Day 7 of 30 Days #BuildingInPublic as an AgriFood Data Analyst. This is the start of my second SQL Analytics Project powered by @DeepnoteHQ. Meet Mkulima Direct FoodMart. It operates in three outlets — Nairobi, Nakuru, Meru — on busy roads, serving working professionals who order mostly on WhatsApp, sometimes online, sometimes walk in. Nakuru and Meru sit in real agricultural zones, so they procure produce, cereals, and value-added goods locally and ship the surplus north. Nairobi has no farmland of its own — it sells the same catalog, sources meat and poultry directly from urban distributors, and leans on transfers from the other two for everything else. I’ve been brought in to answer one question management can’t: is this business actually profitable once every cost is counted, not just the obvious ones? First, I built the data to make that answerable. Every single purchase is its own tracked batch — not a running stock number, an individual event with its own cost, its own supplier, its own fate. From there, it can go one of six ways: sold, rejected on arrival, spoiled in storage before it ever leaves the source outlet, lost in transit on the way to Nairobi, spoiled again after arrival, or left to expire unsold on the shelf. The data includes 14 tables, including over 2,900 batches, 47,700 orders, 82,700 ordered items, and a waste log of 13,900 items. A full year (Jan – Dec 2025) with every kilo accounted for, one way or another. Here’s the uncomfortable part I already found before writing a single line of SQL: total revenue for the year barely clears total cost. And a meaningful chunk of everything purchased never made it to a customer at all. The Question: Is that a pricing problem, a logistics problem, or a waste problem? That’s what the next several days are for. Building the schema, running the joins, and letting the data argue with itself.

The "don't send your data to OpenAI" era is back. This time, it's geopolitical. In 2023, every enterprise security review asked where the data goes. Then the models got too useful, and the questions went quiet for a while. Now the concern is returning to sovereignty. Which jurisdiction sees your data and traces, who controls the model, and whether you can get locked out of the most intelligent model with an executive order. The EU is beginning to fund independence from American tech with tens of millions of euros. Meanwhile, companies both in the EU and the US are starting to look at Chinese OSS models. They’re good, they’re cheap, they’re optimized for high throughput. This makes it very difficult for labs to argue in support of handing over all of your data and outsourcing agentic runs to them fully. I think that companies will start building out their own internal harnesses, with a router switching between closed-source and open-source models for more sensitive or low-effort tasks. Companies like Airbnb, DoorDash, Coinbase, and Lindy are switching to Chinese OSS models. Would you? ____ From Wang, A. H.-E., & Siler-Evans, K. (2026). U.S.-China competition for artificial intelligence markets: Analyzing global use patterns of large language models

Awesome work, congrats! 👏🏻

We've spent a lot of time (years!!!) thinking about the foundational blocks for collaboration on data - I'm glad we've created something that all teams can re-use as they're building agentic workflows within their organization @DeepnoteHQ

Meet Deepnote Agent Workspace: a shared place where people and agents work on the same data tasks and publish them as hosted apps, all grounded in your skills, 100+ native integrations, and permissions. Every major model release gives agents more capability. They can search, write code, query databases, use applications, and sustain increasingly complex work. But capability alone is not enough to scale across the team or reliably reproduce work over time. Most data agents still work inside a temporary chat or a developer’s local harness. They can produce an impressive answer, but they do not automatically inherit the knowledge, permissions, and processes that allow a company to depend on that answer. People do not work that way. A good teammate knows which definitions the company trusts. They know where the relevant data lives and what they are allowed to access. They leave behind work that others can inspect. They learn from corrections. When a task becomes important and recurring, they turn it into a process the organization can rely on. Agents need the same foundation. Today, we’re introducing Deepnote Agent Workspace, a shared place where people and agents work on the same data, using the same organizational context. It brings together the components required to operate data agents across a company: - Skills that capture trusted definitions and procedures, - Agents that perform inspectable and recurring work, - Apps that bring the result into business workflows, - Integrations and permissions that govern access throughout. Interested in trying it out? Reach out, and I’ll hook you up. @DeepnoteHQ

Meet Deepnote Agent Workspace: a shared place where people and agents work on the same data tasks and publish them as hosted apps - all grounded in your skills, 100+ native integrations, and permissions.

Which AI data viz tool makes it hardest to ship a wrong answer? We gave 9 of them the same dataset and the same 4 prompts. Some nailed the math but couldn't show their work. Others produced confident charts built on wrong assumptions. Get the full breakdown with scores, screenshots, and the dataset so you can try it yourself. Link in comments.