← Captain's Log

Anthropic Feature Updates Plus Agent Building Insights

Pencil illustration of a navigator on a modern ship pulling a single chart from a wide cabinet of chart drawers

I’m starting this post with some light feature and product updates, but closing with a super meaty item you might only find interesting if you’re building agents and have coffee in hand. ☕

Claude Duet and the Month of Personal Assistants

First, Claude Duet was announced this week, contributing to what I’d call this month’s AI theme in general: personal assistants. I’m on the waitlist now, but it’s not like I’m chomping at the bit to see this.

I’m already using Claude as a personal assistant. It gives me daily briefs based on access to my calendar, GitHub, email, and Granola notes… and you can do the same without waiting for Duet.

A Quick Note on Instinct

I also continue experimenting with Instinct. But a quick note: of all the personal AI assistants out there now, Instinct is the least documented with respect to the data it collects and uses.

If you had to pick one not to trust, based on that alone, Instinct is it. Proceed (or not) as you will.

Claude Design Moves Into Claude Artifacts

Moving on, Claude Design is being rolled into Claude Artifacts, which also houses Docs, Slides, and Dashboards. I think this makes sense for several reasons, but I’d rather not bore you with that. The point is awareness, and an invite to use something you might not have bothered to look at yet.

Btw, from a pure user interface perspective, some companies are so much better at building the plane while it flies. I think Anthropic is one of them. Who would I say is the worst? Zoom. 😂

🦾 Building Agents With Progressive Disclosure

Finally, I’d like to share some new insights on building agents with “progressive disclosure” in mind. If you like knowing what’s going on under the hood, along with the best practices behind good agent builds, this part is for you.

For the last two years, most of us built AI systems like we didn’t trust the model. We wrote rails (step-by-step sequences with routing instructions), split work across specialist subagents, front-loaded every rule, and wrapped each step in logic so nothing could go sideways.

That made sense when models needed babysitting. They don’t anymore.

The newest models can hold the whole picture, plan their own work, and check themselves along the way. Which means the builder’s job is changing… from scripting every move to designing a clear menu of skills and tools, and letting the model decide what to reach for.

The rails go away… the guardrails stay.

What That Looks Like in Plain Language

The analogy isn’t perfect, but I think you’ll get it. Picture a marketing team handling a client request.

The old way: you have five hired specialists.

One reads the request and decides who handles it. One pulls the data. One writes the copy. One checks brand voice. One formats the deliverable. Each passes a note to the next. By the end, half the context is gone, nobody saw the full picture, and it took twice as long.

The new way: you hire one sharp senior marketer.

On their desk is a binder with a tab for every playbook you run… campaign briefs, brand voice, reporting, launches. They don’t read the whole binder every morning. They glance at the tabs and open the one they need, when they need it.

They also have logins to your systems: CRM, analytics, the asset library. Those logins just pull up the data… the marketer does the thinking. And there are a few house rules they never break, like nothing goes to a client without approval.

Here’s how the pieces map:

  • The binder tabs are progressive disclosure.
  • The logins are thin data tools.
  • The house rules are guardrails.
  • The senior marketer is today’s model.

What Anthropic Found in Real Builds

Anthropic’s article from yesterday shows exactly how this plays out in real enterprise builds:

  • The common approach gives every team its own subagent with an orchestrator on top, which quietly turns your agent into a copy of your org chart. The result is lossy handoffs, a “Frankenprompt” of conflicting rules, and added latency.
  • The new approach is one agent, one context window, with skills and tools loaded on demand. Anthropic cut Claude Code’s system prompt by more than 80% for its advanced models, partly this way.
  • The surprise: in many designs they reviewed, one clear system prompt, 4–5 thin data tools, and firm guardrails covered most of the use case. All that extra structure was doing work the model no longer needs done for it.
  • Subagents still matter for parallel work, verification, or genuine context overload. But the default has flipped: start simple, fold in one subagent at a time, and keep it separate only if your evals say you should.

If you’ve been building Skills, this is your validation. Skills aren’t a workaround… they’re the architecture, the building blocks you must continue to create.

🫡

More From the Log

Pencil illustration of two inspectors with lights examining the hull of a modern research vessel in dry dock

AI OperationsOct 27, 20266 min read

The 20-Minute Security Sweep (A Prompt for the New Reality)

AI just helped uncover 13-year-old flaws in the language behind 70% of the web, and it's helping attackers move faster too. This prompt turns your AI into a personal security coach in about 20 minutes.

Jason CormierCo-Founder