State of AI for development: October 2025

Three things from October that we think change decisions, and what we are telling clients to do about each.

Fast, IDE-tuned models changed the feel of agentic coding

Cursor and Windsurf both shipped fast models built for agentic tasks inside their own IDEs (Composer 1 and SWE-1.5). They are quick enough that you stay in flow rather than dropping into the prompt, wait, prompt rhythm, and they are smart enough for routine work.

What we are telling clients: spend time with one of the fast models to get a feel for what this kind of developer experience does to your working rhythm. Keep planning with a smarter, slower model; hand the routine tasks to the fast one.

Skills: a pattern worth adopting before it spreads

Anthropic shipped skills in Claude Code: a way of packaging a procedure or a piece of domain knowledge as a markdown file plus scripts, loaded lazily from a one-line description. It is far more token-efficient than MCP for this job, and it bundles the things teams keep re-explaining to agents (how to run the tests, how to clone a server). Only Claude Code has it today; we expect the pattern to spread the way MCP did.

What we are telling clients: if Claude Code is your daily driver, start encoding your team’s procedures as skills now. The write-up cost is small and the material transfers if the pattern goes cross-vendor.

AI-native browsers arrived, and deserve caution

OpenAI launched Atlas, its AI-native browser. It is capable, though not dramatically beyond the other agentic browsers, and it shares their problem: significant security risks around what an agent can do with a logged-in session.

What we are telling clients: do not log into sensitive sites inside any agentic browser yet. The capability is interesting; the blast radius of a hijacked session is not worth it.

---

This digest is the public edition. Advisory clients get the full monthly review, a tailored edition for their context and the conversation that goes with it.