The tools are evolving. Has your org?
The agent tooling is maturing quickly; the more durable advantage now comes from the systems organizations build around it.
A year ago, the thing holding agents back was the agents. They lost the thread halfway through a task, couldn't run unattended, couldn't reach half your stack without a config session first. That's not where we are anymore. Long-running tasks, scheduling, automation, a single access layer in front of everything: these are shipping with real harness support now, not living in a demo reel.
Which quietly moves the hard part onto you. The distance between "we have AI tools" and "we've built an org that uses them well" is mostly a function of how you've set things up, and that's the good news, because setup is something you control. Nobody has the finished version of this yet. But the orgs furthest ahead aren't waiting for one, and this week we got two unusually detailed looks at what figuring it out in public actually looks like.
Here's what we've been reading this week.
Lee Robinson breaks down Cursor's Developer Habits Report
Cursor just published its inaugural Developer Habits Report, the most detailed public dataset we've seen on how coding agents are actually changing engineering work, and Lee Robinson's walkthrough is the fast way in. Three numbers that stuck:
- p99 developers ship 46x more lines than the median active user, and merge 15x more PRs than the median PR author. The gap between power users and everyone else is widening.
- Cost per agent request varies nearly 9x across model families, but cost per accepted line only varies ~7x. The pricier models pay for part of themselves by writing code that sticks.
- More than 5x as many agent-generated changes are reaching commits with no manual accept step as at the start of 2026.
Generation is cheap now. The leverage moved to review, architecture, and catching what the agent got wrong and stated confidently anyway. If you're still measuring output in lines of code, you're counting the cheap part.
Basis on building a unified MCP layer
Second week running we're pointing at Basis. They keep publishing the exact writeups orgs need: detailed, honest, trade-offs left in. This one's about Satellite, their MCP gateway sitting in front of 36 tools (HubSpot, Linear, Neon, Slack, and more), giving every agent and teammate one place to connect. Before: the average teammate had 3.2 integrations wired up. After: 17.3.
Integration friction was quietly killing agent adoption, and the fix wasn't a better model. It was one Google SSO login, role-based auth, and centralized logging. Now 95% of the company uses agents, including the people who'd never have configured a single MCP server themselves. They're honest about the trade-offs too: their own wrappers to maintain, a "skills tax" for company-specific behavior. Worth reading in full if you're scaling access past the early adopters.
Claire's no-hype take on Claude Opus 4.8
Anthropic positioned Opus 4.8 as their step-change agent model: 69.2% on SWE-bench Pro, about 15 points ahead of Gemini 3.1. Claire got early access and ran it on real work at high effort. Short version: great on greenfield one-shots (it built a ChatPRD prototyping feature start to finish, about 20 minutes autonomous, worked first try), but it struggles through the last 10% and hallucinated more than expected, confidently reporting whether it had checked GitHub, validated a bug, or looked at real data when it hadn't. Opus 4.7 beat it on strategy work.
Use it for greenfield prototypes. And double-check it exactly where it sounds most sure of itself. That's where it's most likely making things up.
One thing to try this week
If you don't have any data-driven insight into AI adoption in your org, fix that this week. Crude measures are fine, perfection is not the goal. You don't need to build a dashboard, just try to learn something.
Could be token consumption, or PR count, or whatever you can easily and reliably-ish get your hands on. Use an agent (good forcing function for you!) and take the first step. Our bet is you'll find out more than you bargained for.
Have a great weekend. Spend time outdoors. Build something fun.
— Claire + Zach