The Agent Anthropic Hasn't Built Yet: Twenty-One Frictions From One Upgrade

This is my longest blog post ever. And I am typing really loudly.

I work with 12 AI agents that I designed to provide value in different areas of my business. They share general context, and they communicate with each other, but each has developed unique skills related to its role. They have mechanisms to ensure that they continuously improve over time. I have given them names. I used Claude Projects to identify and customize each agent. On Sunday, Anthropic changed how my Claude Projects function.

At first I was immensely annoyed. My agents changed overnight. Imagine an office where people came in one morning and suddenly started operating with very different behaviors. That's probably a sitcom episode. I was annoyed all day while I tried to understand the changes. Jump to the end. The new Claude Projects design is a vast improvement. Good for them. My takeaway is much more focused and nuanced than my initial spew of profanities. Change can be good. Poorly documented change is not good.

The AI industry is moving faster than I have seen any technology move, and the pace is still accelerating. The labs now say so themselves. Anthropic reported that more than 80% of the code merged into its codebase as of May 2026 was written by Claude, in a paper it titled "When AI builds itself" (Anthropic; The Next Web). By November, it was warning that AI that improves itself could arrive sooner than expected (Android Headlines). AI is now doing much of the work of building AI.

I have always been in favor of progress and change. But I think Anthropic in particular (the company I work with most) is missing a major opportunity, and I have two observations.

First, the Claude agents themselves are rarely up to date on their own platform. Mine gave me directions to settings that had moved, and told me a setting worked one way when the docs say another.

Second, Anthropic could use a specialized Claude agent to help users adapt to releases that now arrive this fast. On Sunday, November 15, it took me four hours to work out what had changed in the new release, how it affects what I do, and how to adapt. The missing piece is a helper agent that knows what is new, learns what you do, and then helps you adapt. Without one, I did it all on my own, and that is not how AI is supposed to work.

Here is the evidence: twenty-one frictions, twenty of them hit on that Sunday while restarting Sam, our marketing agent, and Clarice, our chief of staff, under the new setup, and one two days later, with the fix for each. None of them was a bug. Each one was a safety default or a change in how the pieces fit, and each one cost me some minutes and some patience. The rest of the post will be useful to anyone using Claude when they get the new Projects upgrade, to Anthropic so they can fix this problem, and to me as I design a new agent to help other Claude users and me learn Claude at the same pace new features are being released. Skip to my Final Thoughts.

Where the agent's memory went

1. A restart is a clean slate. I changed the project instructions, and the agent started a new thread with none of the earlier context. That is the right design: the agent's memory lives in our files (a folder we call Start), not in a chat. That means the first minutes of any restart go to rereading 12 files before the agent can do anything useful.

2. The agent talks about itself in the third person. The new setup puts a coordinating layer between the agent and me, and that layer reports on it: "Sam is writing the draft." Clarice did the same: before the move she was the one talking to me; after it, the chat described her, as if the agent and the conversation had split apart. I hire an agent so it can work like a member of staff, and staff says "I." I asked for first person, and I'll keep asking until it is the default.

Where the files are

3. Read is not write. The Google Drive connector could read every file in Start, but almost none were writable. The connector was signed in as my personal Google account, and the folder belongs to the company account. One fix is to share the folder with that account as Editor. The fix I used: give the agent the copy of Start that Google Drive syncs to my PC.

4. Every new folder asks first. Giving the agent that local folder put an Allow prompt in front of me. One click, and it holds for the session. I'd keep this default.

5. Delete is off until you turn it on. The agent tested its write access by creating a scratch file and then removing it. The create worked. The remove failed because deleting in a connected folder is off by default, so our company folder now holds an empty file named .write-test. I agree with this default. I want the agent to test with something it can leave behind on purpose.

Where the web is

6. The cloud sandbox has a guest list. I sent the agent a link to an archived 2000-era InfoWorld page. The cloud environment's network is allow-listed by default, and web.archive.org was not on the list. The default level is called Trusted. The fix is to raise it to Full (items 9 and 10 cover what that took), or to run the agent on my own PC through Remote Control.

7. A second gate sits on the PC. When the agent retried the same page from the shell on my PC, an automatic safety check stopped the request. Two independent guards on one fetch is defensible. It is also why the agent then asked me to paste the page, and I told it, in words I won't print, to look it up itself. It did. Two other public sources had the answer. And even with the network fully open, the Internet Archive itself turns away automated readers: the cloud fetcher, a plain download, and the proxy on my PC were all refused. That door was never going to open for a robot.

8. Claude in Chrome now asks twice. Until this update, the agent used my own Chrome, through the Claude in Chrome extension, without a separate yes. Now it takes two steps. First, I write the approval in my own words in the thread; a yes relayed by another agent does not count. Then I click Allow for each site in Chrome, and that prompt can time out while I am in another window. The intent is sound: an agent in my signed-in browser can act as me. The cost is that a two-minute errand now needs me at the keyboard.

Where the settings are

9. The network switch is hard to find. Network access is not under Settings. It lives in the environment picker you see when you start a new session at claude.ai/code: hover the environment's row, then click the gear. The first directions I got pointed somewhere else, and that cost a round trip.

10. A new setting may need a new thread. When I switched Network access to Full, the session that was already running did not pick it up while I waited. Anthropic's docs say the change should reach running sessions within about a minute; in my case, only a new thread got it. Project instructions do work this way, which is why item 1 happened at all: a change to a project's instructions reaches only new threads, never the coordinator that is already running.

11. No built-in browser on the PC. The agent's thread had no built-in browser on my machine, so the only browser left was my own Chrome, with the two approvals in item 8.

Where the projects are

12. Archived projects are hard to find. I archived the old projects during the move. In the desktop app I found no clear path back to them.

13. The instructions moved house. Each agent's project instructions seemed to vanish. They hadn't: they now live in the project on claude.ai, not where I had set them up. Nothing was lost, but for ten minutes I thought it was.

14. Approval prompts show up on another screen. When Clarice asked to open a second folder, the Allow prompt appeared in the Claude app on my PC, not in the chat where I was talking to her. I told her I didn't see it. Meanwhile she was waiting on it.

15. One folder per approval. The project came with one approved folder. The old archive Clarice needed sat right next to it in Google Drive, and it needed a fresh approval of its own (the prompt from item 14). I'd keep the default; I'd put the prompt where I can see it.

16. The load marker didn't survive. Our agents check for a line in their instructions that says the project loaded correctly, and stop if it is missing. The migration dropped that line. The agent carried on anyway, because the rest of the instructions were right, but the safety check we built was gone.

17. A restart drops a pending approval. While a folder prompt was still waiting for me, the agent's session restarted in the background. The request was closed unanswered, and the agent had to ask again.

18. A scheduled job went quiet. Clarice's daily sweep of my email and calendar was approved the day before the switch. Her log shows no run since. Moving the agent didn't move its schedule.

19. A scheduled task's instructions are signed to the PC. Our Medium task stalled on posts that lost a table on import, and I ruled to ship them as-is. The agent tried to update the task's instructions to match and was refused: a task that runs on my PC takes new instructions only when I approve them in a chat on that same PC. So a one-line rule change waits until I am at that machine. A sound guard against a hijacked prompt, and one more errand for me.

20. An approval doesn't travel between threads. When an agent works on my PC through a separate Remote Control thread, a yes I already gave in the project chat doesn't count there. I have to type it again inside that thread. That holds even after I made "never ask me to do something you can do" our top rule, because only I can type my own approval.

21. A habit that didn't migrate. Two days later I asked for an email draft to a colleague. In the old project, that meant a draft waiting in my Gmail, ready to edit and send. Under the new setup, the work ran in a background thread, and the draft landed on an in-app review card inside that thread. I never saw the thread, so I never saw the card. I had to ask where the draft was, have it redone in Gmail, and then add a line to the project instructions: email drafts go to Gmail, never to a card. A review card is a fair default for someone starting fresh. But I wasn't starting fresh. "Drafts go to Gmail" was never written down anywhere; it was a process/habit the old project had settled into, and the upgrade didn't carry habits. An upgrade changes defaults without telling you, and a preference that lives in habit rather than in a written rule is the first thing it loses. Write your preferences down before the upgrade, not after.

What's better

To be fair to the update, one change is a clear gain. A project now runs its own work as threads, and each thread shows its own step list while it works. I can see what an agent is doing, what is done and what is next, without asking. When an agent needs a call from me, it puts a decision card in front of me with options and a recommendation, and I answer with one tap or one letter. This morning I settled the Medium question by typing "A." That is the shape I want for managing agents: they run the steps, and I make the calls.

What Anthropic said it would do, and what others ran into

I wanted to know whether my Sunday was typical, so I had Sam, our marketing agent, research the release and the reaction to it. Here is what the record shows, as of November 15, 2026.

There is no single named update. The core of it is "Projects" in Claude Code, a public beta that Anthropic announced on September 17, 2026, in a post titled "Projects redesigned: from folder to conversation" (Claude blog). It opened first to some Pro and Max plans; Team and Enterprise came later. A second wave on November 13, 2026, made the coordinator "much more proactive" (@ClaudeDevs on X), and third-party write-ups say it added thread reminders, /btw, /fork, and read-only sharing (explainx.ai, dev.to). Neither date appears in Anthropic's support release notes. So the "November update" I hit is the second wave of a September launch.

What it was for. Anthropic's pitch is that a project stops being a folder of files and becomes an ongoing conversation. In its words, "Brief Claude in the project the way you'd brief a chief of staff," and "you describe what needs to get done, and Claude manages the work" (Claude blog). The coordinator splits work into parallel cloud threads that keep running after you step away, with shared memory and a library of files (Projects docs). VentureBeat summed up the design as "threads do the work, Claude directs it" (VentureBeat, Sept. 17). That is the staffing-agency model in software: a manager who delegates, and staff who report back.

Most of the other pieces are older. Remote Control dates to about February 2026, scheduled tasks (formerly "routines") to April, and Claude in Chrome to August 2025. Projects now drives them. Cloud environments offer four network levels: None, Trusted (the default), Full, and Custom (cloud environments docs), though some of Anthropic's own pages call the default "Limited." The permission tightening I felt arrived in small steps: Chrome site-permission changes in Claude Code releases from late August to mid-October, and on November 13 a release that made scheduled tasks and sandboxed commands ask more often and narrowed "Always allow" (Claude Code changelog). None of it was announced as one change, which is why it landed on me as twenty-one.

One correction to my own log. The docs say a change to network access reaches running sessions within about a minute (cloud environments docs). That is not what I saw in item 10, but it means the rule is "should apply," not "only new threads." Environment variables are different: those do wait for a restart. I found no official documentation for decision cards, for deletes being off in connected folders, or for the rule that a Chrome approval must be in my own words; those show up only in reports from people who hit them.

What people praised. The main Hacker News thread on the November wave drew 99 points and 95 comments (HN). The heaviest multi-agent users were glad to retire their homemade setups: "I'm excited to have a version of this I don't have to maintain myself anymore" (HN), and "So much cognitive overhead just managing that, this is so much better" (HN). People liked threads grouped by status and running a session from a phone through Remote Control. Boris Cherny of the Claude Code team wrote that Projects "are how I write pretty much all my code these days" (HN).

What people complained about, roughly in order of how often:

  • Usage burn. One developer asked a simple question, and it "spawned 7 sub-agents and ran me completely out of usage in about 10 minutes" (HN). Anthropic's docs warn that projects use plan limits faster, and VentureBeat flagged the same risk at launch.
  • Approvals that don't carry over or don't stick. One thread refused a save "because the approval reached me through the coordinator, not from you in this thread" (HN). That is my item 20, almost word for word. On GitHub, people ask for a folder "Allow" that is remembered (#101391), a default "skip all approvals" mode (#98159), and delete rights on a subfolder so git stops leaving lock files behind (#100962), which is my item 5.
  • Not seeing what threads are doing. "If you want to know what they're doing you're supposed to ask your coordinator. That's crazyville" (HN). The step lists I praised above answer this, and not everyone has found them yet.
  • Network settings that don't take. Several GitHub reports say a session ignored "All domains" or the allowlist (#93512), and another says the Chrome extension blocks reddit.com outright (#95326).
  • Projects lost in the move. Reports say the upgrade forced classic projects into the archive (#101369), my item 12, and that a recycled coordinator lost its decision-card tools (#104059).
  • The rollout itself. Desktop only, no command-line version yet, and a push into Anthropic's cloud that some read as lock-in. People want threads that run locally (#99156).

On X. Anthropic's November 13 announcement drew about 18,000 likes, 620 replies and 1.5 million views in two days (@ClaudeDevs). Sam read roughly 150 of the replies. The most-liked was about access, not features: "Whose door do I have to kick down to get access to this on my team plan?" (X), echoed by Enterprise customers who still don't have it. Next came the same friction I logged: "Projects is amazing guys really. But you have to loosen up by the safety check and safety blockers. Or allow us to use bypass on projects" (X). Usage came up again: "It nukes usage" (X). One person asked when the built-in browser comes back to projects (X), my item 11. Another said the network-settings button can't be reached with a screen reader, so they use a second AI agent just to open it (X), a sharper version of my item 9. The angriest reply was about the migration itself: "my whole workflow is broken now," after the desktop app "updated itself in front of me without asking" (X). On the other side, people who had built this by hand were pleased: "I've run something close to it by hand for a year: one coordinator, parallel sessions, background jobs, reminders" (X), and "Incredible updates. Been getting so much done with Claude Projects!" (X). Commentator Martin Szerment summed up the economics: the coordinator "costs you almost nothing. What you pay for is every thread it decides to spin up" (X).

How Anthropic responded. Cherny answered in the Hacker News thread: he asked for bug reports on the usage case, pointed people to the Remote Control setting for local access, and said a command-line version is in the works and the Desktop rollout is gradual (HN).

The competition moved first. Cursor shipped its own coordinator-based Projects about a week before Anthropic and made it multiplayer on November 10 (Cursor). One Hacker News commenter noted Cursor had "the exact same concept and names."

My read (subject to change). Anthropic built what I asked for in item 2 and in "What's better": a manager layer that delegates and reports. The friction is the cost of bolting that manager onto older permission systems, each designed for one human at one keyboard. A dev.to reviewer put it well: "More threads means more output, not more review capacity" (dev.to). Someone still has to approve, review, and own the work. That is the job we sell.

A note on sources: this covers Hacker News, X, GitHub issues, Anthropic's own docs and press. Reddit was out of reach: the Claude in Chrome extension blocks reddit.com outright, the same block reported in GitHub issue #95326, so the agent could not read it even with my approval. GitHub issues are summarized from titles and reaction counts.

What it adds up to

Count them: twenty-one frictions, about 45 minutes per agent, zero lost work. Five are safety defaults I'd keep (items 4 through 7, and 15). Four are identity and wiring problems that a one-time setup fixes (1, 3, 13 and 16). Eight are product choices I'd change: the third person, the hidden network switch, settings that skip running sessions, no built-in browser, hidden archives, prompts on the wrong screen, restarts that drop approvals, and approvals that don't travel between threads (20). The Chrome double approval and the PC-signed task instructions (19) sit in between: right in principle, too heavy for a routine read. And two were ours: a scheduled job that has to move with the agent (18), and a preference we never wrote down (21).

Current view (subject to change): each layer of permission makes sense alone. Stacked, they turn a capable agent into one that asks a lot of questions on day one. What I want is a setup checklist that clears all of them in one sitting, so day two runs without them. That is the checklist we want for our own clients' installs.

Final Thoughts

The work of managing an agent is mostly this kind of plumbing: where its memory lives and what it is allowed to touch. None of it is glamorous, and it all decides whether the agent saves you time on day one or on day ten. Which brings me back to the opportunity. Every one of these twenty-one items is something an agent could have told me before I hit it: what changed, which parts of my setup it touches, and the fix. I would still rather Anthropic shipped that agent for everyone.

In the meantime, here is what we are building at Coworkers.Global. Chet is our managed agent for Claude itself. His job is to be up to date on everything Claude: models, plans, features, connectors, plugins, skills, and the settings that decide what an agent may touch. He works from a knowledge base we keep current rather than from what a model remembers from training, and every answer he gives carries the date it was last checked. That fixes my first observation. For a new user, Chet sets up the shared folder your agents work from, explains how Claude works in plain language, and offers a short tutorial when you seem stuck. He will not have every answer on the day a release lands. He is built never to guess, and to say when he needs to check. (Coworkers.Global is not affiliated with Anthropic.)

If you are running agents on Claude and hit a friction I missed, tell me which one.

Regards,

Charles Stack
Founder, Coworkers.Global

Coworkers.Global is an AI staffing agency. We place managed agents into organizations that need dedicated expert knowledge work. A managed agent is an AI specialist provisioned for a specific role, trained on your context, supervised by a person, and accountable for its output. The first, Alex, evaluates startup business plans for fundability, informed by human expertise and research, and calibrated against real investor decisions. We are early-stage with paying clients, and we still lead with the quality of our judgment rather than customer logos. The managed AI staffing agency.
Get new posts by email

What we're learning building a startup with managed agents, plus notes on raising. Weekly newsletter, no spam.

Want more out of Claude today? Chet is available now, $99 a month. Every other role is on the waitlist.