The frontier moves every week.We test it for you.
An independent, curated view on models, tools, protocols and agents: what is real, what matters for your setup, and what to ignore.
We're building Frontier Watch with a small group of AI leads. It starts with a 30-minute conversation.
| What moved | On our bench | Call |
|---|---|---|
| New open-weights coding model | Close on our evals, a third of the cost | Benchmark it |
| MCP servers for your SaaS stack | Auth still uneven | Pilot, guarded |
| Always-on personal agents | Ran a week on real tasks | Watch |
| Yet another eval platform | Overlaps what you have | Ignore |
Every update has three layers.
From what moved in the field to what it means for your stack. Click through an example.
- Agent protocols are settling. MCP is becoming the default way to connect agents to tools and data.
- Coding agents run for hours. Long-running agents now finish multi-step tasks without a person at every step.
- Open-weights models closed the gap again. For many tasks, self-hosting is a real option.
- Integration beats model choice
- Evals become your moat
- Three coding agents on the same five tasks from a real codebase: time, cost and how often a person had to step in.
- Always-on personal agents for a week on real work: where they helped, where they made things up.
- MCP servers for common SaaS tools: what works, and where authentication breaks.
- Results with the setup we used
- Repeatable on your own code
- Pilot one coding agent with two teams, on refactoring work. Review in six weeks.
- Wait on always-on agents until access controls mature.
- Write an MCP policy before teams connect their own servers.
- Depth for your team
- Decisions for leadership
Illustrative example. Your note is written for your setup.
We're building it with the people who'll use it.
Frontier Watch is new. We're starting with a small group of AI leads who help shape what goes into it.
- 130 minutes
A conversation
What you follow today, what costs you time, and what you'd want from us.
- 2First editions
Founding member
You get the first updates and tell us what to keep, cut or add.
- 3Every month
The update
The field, what we tested, and what it means for your stack.
- 4Every quarter
Peer session
AI leads from different companies, one room, one guest who shares their practice.
No commitment to start. The conversation already helps us, and we hope you too.
- You lead AI, data or engineering: head of AI, data lead, CTO
- Your teams build with models and agents, or are about to
- You want an independent view from people who test, not sell
- You want a general newsletter about AI
- You're looking for someone to build your tools
- Your people are just getting started: then start with the Academy
Practitioners, not a platform.
Ben Mellaerts
Strategy, change and adoptionLeads AI adoption programs in big organizations. Speaker and podcast host.
Niels Rogge
Open models and AI researchOne of Hugging Face's early engineers. Reads new research daily and teaches it in plain language.
Tim Leers
Agentic AI products and platformsBuilt the agentic AI practice at dataroots and Talan. 50+ deployments, millions of users.
Questions.
How is this different from a newsletter?
We test what we write about, across many companies. Every update ends in three calls for your setup.
Do you use AI to make it?
Yes. Agents follow the field. People test, filter and write.
What does a founding member do?
Talk to us for half an hour, read the first editions and tell us what's useful. We set prices together with our founding members.