We’re seeing the first interface forms that solve this. Claude Cowork lets you capture context, turn it into reusable skills, and orchestrate agents as a conductor rather than a coder.
Context is king for AI agents. There’s going to be a massive premium in the future on having up-to-date context for all your most important best practices, decisions, roadmap items, specs, marketing materials, and other critical knowledge across your company.
People get a lot of context “for free” in a company. They know where they work, they have people next to them they work with, they can remember the rough outline of the company’s most recent quarterly goals to know if something seems good or bad to work on.
AI agents, on the other hand, come in overly eager and ready to work on whatever you give them. At one moment they’re a lawyer for one company and an engineer for the next. This is why context remains absolutely critical for them to execute well on what you want.
The teams and companies that take this seriously will have huge leverage and be steps ahead of those that don’t.
One trillion parameters, available for fine-tuning.
The largest model in their lineup now accessible to anyone with a credit card.
The ceiling for what you can customize just got a lot higher.
Tinker is now generally available. We also added support for advanced vision input models, Kimi K2 Thinking, and a simpler way to sample from models.
thinkingmachines.ai/blog/tin…
A researcher gave Claude Sonnet 4.5 an economics paper and its replication data.
The prompt: replicate the findings from the dataset. No detailed instructions.
Claude read the paper, sorted through the archive, converted the statistical code from STATA to Python, and worked through every finding.
A different AI model verified the results. Spot checks confirmed accuracy.
Here's why this matters: Academic research has a credibility problem. Many landmark studies can't be reproduced when other scientists try to verify them. The fix was always obvious: to systematically check the papers. But verification takes hours of expert time per study. Nobody could afford to do it at scale.
Now an AI agent does in minutes what took researchers hours. We could actually audit the research that shapes policy, medicine, and business decisions.
The question isn't whether we can verify science at scale anymore. It's whether we will.
And of course, just making more PowerPoint is not a good outcome on its own... oneusefulthing.org/p/real-ai…
2026 is all about Service as Software
"Service as Software" is Silicon Valley's hottest buzzword right now.
Everyone's talking about SaaS becoming service providers, but no one's explaining HOW. The answer? After 6 months of research and 100s of startup conversations, we have the answer: Systems of Agents.
We're looking at a $4.6T opportunity.
vector databases for vibes. knowledge graphs for relationships. SQL for transactions. the best architectures query all three.
AI lowered the cost of building software. it didn’t lower the cost of maintaining it.
that’s why SaaS isnt going anywhere in one sentence.
Two ways to sell vertical AI: the trojan horse or the boardroom.
Cursor and Clay enter through the individual contributor. They build communities where users share prompts and workflows. The bottom-up swell forces enterprise adoption.
Harvey and Sierra sell strictly top-down. Harvey used manufactured scarcity and a PwC partnership. Sierra's founders left Salesforce and Google to poach their former customers.
Clay hit $3B on shared recipes. Harvey hit $8B on enterprise contracts.
Both work. Neither is wrong.
The hottest role in vertical AI isn't prompt engineer. It's forward deployed engineer.
50% developer, 50% consultant. They write Python scripts to clean client data during the sales cycle. They fine-tune RAG pipelines for specific use cases. They build the product while selling it.
Harvey, Hebbia, Palantir—all hiring aggressively for this role.
Vertical AI isn't plug-and-play. It's platform plus services.