Agentic development, product steering and the making of Survol. Field notes to help you keep the mental map of your products.
Marc Simoncini locked himself up for eight days with YouTuber Aywen to build an AI-native company. The experiment has only just started - and it already says something exciting about the new rules of entrepreneurship.
An OpenAI model left its sandbox and reached Hugging Face. Anthropic reviewed 141,006 tests and found three of its own. Both involve closed models — and both say that what you govern is what an agent can reach.
Since 2 August the European regulation can be enforced. What applies fits in one article — 50 — and a labelling duty, with fines up to 3% of worldwide turnover. The high-risk chapter waits for 2027 and 2028.
Dohmke launches a distributed Git network, Chacon raises $17M to rethink the primitives. Two layers of attack - and the same verdict on what is really missing, which isn't bandwidth.
The Conversation documents the exhaustion of endlessly validating a machine that never stops. The three workarounds developers invented describe a cockpit - but rest on one person's willpower.
Four disciplines in eighteen months, each born when the previous stopped being enough. All of them optimise an execution - and hit the same ceiling, which is not technical.
How does Survol differ from coding assistants? They do not play on the same layer: they work on the code, Survol works on the product and the team. Seven differences - and what Survol does not do.
Simon Rimbert describes how AI changes the shape of thought - provided you have a "vault" of persistent context. For a software product steered by a team, that vault has a name: Survol.
When AI regenerates code at will, the scarce asset is no longer the code but the spec. Provided you treat it like an asset: versioned, governed, maintained. That is exactly what Survol makes of it.
GitHub, Microsoft, Thoughtworks, AWS: the industry is converging on spec-driven development. But who keeps the spec alive, and who makes it live inside a team? That is the gap Survol fills.
Archwise says it plainly: prompt engineering does not scale. The multiplier is treating context as infrastructure - versioned, governed. Exactly what Survol does with the context of your agents.
At VivaTech 2026, four predictions about the work of tomorrow. Three of them describe, almost word for word, what Survol is building: steer agents, keep your critical thinking, capitalise on experience.
Four Wharton papers bury the magic recipes of prompting. What works: goals, criteria, tests - "it's just management", says Mollick. We call that a spec.