I refused to believe Claude Code got worse. It had.

I went quiet on April 7. I came back today. Here is what happened in between, in roughly the order I reacted to it.

Mythos Preview shocked me. I still do not know what to do about it.

Anthropic announced a model so good at finding security bugs they decided not to release it. Project Glasswing now runs it, invitation only, with most of big tech. I read the post twice and noted it. No take. Some news is just news.

Claude Code felt off. I blamed myself for two weeks.

Sometime mid-April, sessions started feeling slightly worse. More verbose. Skipping context I was sure I had given. I told myself I was tired and kept working.

On April 20 Anthropic published an apology. Three separate changes, three separate quality regressions, all rolled back. Usage limits reset for subscribers. My gut had been right and I had spent two weeks ignoring it.

Lesson I want to remember: when the tool feels different, the tool is probably different. Check the changelog before blaming yourself.

Claude Design: three design systems, one actual win.

Claude Design launched April 17. It has its own weekly allowance, separate from chat and Claude Code. I burned through mine in about a week.

Attempt one: I pointed it at fieldnotes-ai.com. Got a clean design system for a site that already exists in code. Cool. Useless.

Attempt two: I pointed it at Forma 36, Contentful's open-source design system. Great output for someone building a custom Contentful App. I am not.

Attempt three, the one that worked: I pointed it at the deck templates we use at work. Now I describe the slides I need and it produces on-brand decks in about half the time. I build decks a lot. Cutting that in half is the most concrete productivity win I have had from any AI tool this year.

One thing worth flagging: I hit my weekly limit after three design systems and a handful of slides. The system-building tasks seem to chew through allowance faster than the actual slide generation. So if you are evaluating Claude Design, I would skip the "build me a design system" experiments and go straight to making the deliverable. The allowance will go further.

Managed Agents: still building, slower than I want.

I started a multi-day session on the Research Agent. I made progress. I also kept getting distracted by every new release, which is a humble way of saying I read more than I shipped. Partway through a build that should already be done.

The good news: memory for Managed Agents went into public beta on the same managed-agents-2026-04-01 header. The Research Agent needs persistent memory to do its job. I had been quietly dreading building that layer myself. It now exists. Significant unblock.

April 19: Vercel got hacked.

The attack came in through Context.ai, a third-party AI tool used by a Vercel employee. OAuth into Google Workspace, pivot into Vercel, decrypt non-sensitive environment variables.

fieldnotes-ai.com runs on Vercel. I rotated everything that was not marked sensitive: Contentful management token, Supabase service-role key, Resend, Voyage AI. Flagged the new ones as sensitive going forward. Activity log was clean.

The part that stuck: an AI side tool was the way in. Every OAuth grant is a potential pivot point. I am going to be more careful about what I connect to my Google Workspace from now on.

What the pause actually was

I called it overwhelm on April 7. By April 26 it had turned into something more useful: one workflow that is genuinely faster, one security cleanup I had been lazy about, one regression caught, one unblock I needed, and a multi-day Managed Agents session in progress.

Not lost time. Just unwritten time. Next note will have actual code.