The Future of Work with Claude

Study Guide

Overview

In this conversation from the Claude YouTube channel, Boris Cherny (Head of Claude Code) and Cat Wu (Head of Product, Claude Code) walk through how AI-assisted work has evolved over the last few years and introduce Claude Tag, a new way of working where Claude proactively jumps into team channels like Slack, does work that can span days or weeks, follows up on its own, and remembers what you told it for next time. They cover the research that made it possible (long-horizon task performance and memory), the emergent usage patterns inside Anthropic and among early customers, and the striking internal stat that roughly 65% of the product org's PRs are now written by Tag, and climbing.

Key Takeaways

  • Two leaps in two years. Boris traces the arc from type-ahead (deciding each line yourself) to agents writing whole functions, files, and features, to Tag doing "the whole thing," including entire experiments end to end and his data analyses.
  • Claude Tag is proactive, not reactive. Instead of opening Claude and asking it something, you add Claude to a channel. It knows when to jump in, does the work even if it takes days or weeks, follows up, and remembers instructions for next time.
  • Long-horizon research is the enabler. On the latest METR evals, the models can work 16 hours at a time, and are now in a zone where the team can't even accurately detect how long they can work. Tag can also self-schedule follow-ups after days, weeks, or months.
  • Memory finally works. The model sets reminders for itself and remembers instructions from all users over time. Boris says the team tried to get memory right for Claude Code for years and "it feels like we finally got it right."
  • It's multiplayer. Most AI tools today are one person plus Claude, then copy-pasting output to the team. Tag brings Claude into the middle of the work so the whole team can guide a session to a better result.
  • Adoption spread by observation. Because Tag works in public channels, everyone can watch how expert users leverage it and carry those patterns into their own projects — a diffusion of best practices Cat calls novel among AI tools.

From Type-Ahead to Tag

The video opens with a joke that lands as a thesis: "What do you do if there's no Tag in the chat?" "I don't know. I can't remember a world without Tag in the channel." Asked to rewind three years, Boris recalls that two years ago he was still using type-ahead — "that was what AI could do as an engineer." He was still deciding what line of code to write, with the AI helping a little. Then agents began writing whole functions, whole files, whole features. Now, with Tag, "it just does the whole thing": it can write the feature, run an entire experiment end to end, and handle his data analyses. He calls it "two leaps in two years."

Cat frames the same arc as a transition in who drives: from one person in the loop typing a line at a time, to one person sitting with ten Claudes typing a bunch of features at a time, to a point where Claude is actually driving a bit more, and it's not just one person anymore — an entire team interacts with it.

What Is Claude Tag?

In the past you had to open Claude, ask it something, and Claude would do the work using your tools and your computer. That was a reactive experience. With Claude Tag, Claude jumps in proactively. It knows when to jump in, it will do the work even if it takes days or weeks, and it will follow up. The part Cat highlights as coolest: "it'll remember what I told it for next time." Instead of prompting Claude every time and having to remember to do so, you just add Claude to a channel. It proactively does the work, everyone gets to see it and participate because it's multiplayer, and it remembers for next time.

The Research Behind It: Long-Horizon Work

Asked what on the research side enabled Tag, the answer is years of work making the models more long-running and autonomous. On the latest METR evals, the latest models can work for 16 hours at a time, and they are now in a zone where the team "can't even accurately detect how long it's able to work for." In Claude Tag, Claude can self-schedule work for itself, so the effective horizon actually increases over time: Claude can schedule a follow-up after days, weeks, or months. Tasks the team couldn't trust it with before now succeed "almost every single time."

Boris says he personally has a bunch of Tag sessions running for days or weeks — he thinks one may be about a month old. It's a long-running experiment: Tag checks in every day, checks the data once in a while, sends a bug fix when a bug needs fixing, and he just sees the pull requests coming in and gets data readouts every day.

Memory and Personality

The other big innovation is memory in the model. The model can set reminders for itself and remembers all the instructions all the users have given it over time. Boris: "Memory took a long time to crack. I think we tried to get it right for Claude Code for years. And it feels like we finally got it right. It feels really good to use." In practice, if you tell Tag in a channel to monitor only for one type of issue and not other categories, it will remember that for that channel forever after — and if someone else later expands its scope, it adjusts.

People also love Tag's personality and EQ. To the common worry — "how do I know it won't jump in on every thread and be annoying?" — the answer is that Tag is trained to have a good sense for when it's needed and can take a back seat. If it ever goes too far in the wrong direction, you can tell it to jump in less or jump in more, and it will remember that for the future.

Usage Patterns and Emergent Behaviors

  • Customers are figuring it out. Cat says the coolest thing has been watching customers discover uses: it can write your PRs, debug production issues, and do data analysis — all you have to do is hook it up to your tools. That's only possible "because the memory is just so good."
  • Emergent channel rituals. Boris describes telling Tag that in a Q&A channel every answered question must be checked off — and it just does it, answering questions and reacting with a check every time. Anthropic has a data-questions channel where Tag now answers all the questions, and a Claude Code feedback channel where "Tag just fixes everything now."
  • From tagging to standing instructions. Cat's most common data-question workflow was posting in the data channel and tagging Claude Tag for a solid first pass. She kept doing this across channels, got tired of tagging it, and just told it "please respond every time" — a demonstration of how powerful the memory is. "You can just tell it, hey, remember to always do this thing, and it just has your back."
  • Learning by observation. Because Tag works in public channels, everyone can observe how expert users leverage it and bring those patterns to new projects. Both internally and with customers, they've seen a diffusion of best practices that Cat considers very novel for AI tools. Boris was surprised how fast it took over: Tag is now in every feedback channel and every data channel; it started with a few people, and others picked it up quickly once they saw it in use.

Proactive and Multiplayer

Cat contrasts Tag with existing tools like Claude chat, Cowork, and Claude Code: you typically have to remember to open them, so it's a more reactive experience. With Tag you give a higher-level objective — "put up PRs for every bug in this channel" — and it puts up a PR for every single one without reminders. The second big shift is that it's multiplayer: rather than you and Claude working alone and copy-pasting output to your team, Tag brings Claude right into the middle of your work so multiple people can guide the session to a better output. Several dynamics compound: everyone learns to use Tag better by observing others, Tag learns to work better with the team by hearing their guidance, and the customer gets a better result because the whole team can jump in to nudge a PR to its best possible state.

Boris loves this for a different reason: he wants everyone to contribute to the codebase, but many people are afraid because they'd have to open a terminal or the desktop app and deal with git and code checkout. "With Tag, they can just do it."

Public Channels, Self-Serve Answers, and the 65% Stat

Cat stresses the importance of your collaboration platform having public channels so Tag can monitor the projects you're tracking. As a PM often working on five to ten features, she has Tag look at the status of every single one and give her a daily report. Another big unlock is self-serve questions: when people onboard to the company, instead of asking legal whether they can say something or asking HR what the benefits are, they just tag Claude. Because Tag is connected to the company's source-of-truth files, it gives a fast answer no matter what time they ask.

The single biggest thing, per Boris, is how much more productive it has made everyone, driven by Anthropic's internal version of Tag. In the product org — the part of the company where they sit — the share of PRs written by Tag is around 65% and climbing. He thought Claude Code was the thing that made engineers faster; Tag makes them faster still, largely because of the interaction pattern: you ask Tag to do something, it responds, and you've already jumped to the next task. The goal is to set things up so Tag can do the work without you checking in all the time.

When to Use Tag vs. Claude Code

Boris runs hundreds of Tags but remains a Claude Code power user. He used to use Claude Code for everything, then started using Tag for simple fixes — a button off by a few pixels, a simple data question. Over the last few weeks, as he's grown more comfortable, he uses Tag for more and more, including more complicated work. It can verify its own work because it runs in the same remote sandbox used for mobile and the desktop app and uses the same agent SDK — "so it's just as intelligent." Layer memory on top and each channel can carry its own verification preferences: often Tag fixes a bug and then posts a video back in Slack, so he never has to leave the channel.

What's Next

Claude Tag launched in Slack, and the team is excited to bring it to more collaboration platforms like Microsoft Teams — the goal is for every knowledge worker to be within arm's length of Claude, no matter where their work happens. Cat's closing point: Tag isn't just changing how individuals work; it puts Claude in the center of teams and transforms entire orgs. It's built to be incredibly customizable, just like Claude Code, and she's excited to see how organizations make it their own.

Notable Quotes & Data Points

  • "I can't remember a world without Tag in the channel."
  • Boris: "It's been like two leaps in two years."
  • "With Claude Tag, Claude jumps in. It's proactive. It knows when to jump in. It'll do the work even if it takes days or weeks. And it'll follow up."
  • METR evals: the latest models can work 16 hours at a time — now in a zone where the team can't accurately detect how long they can work.
  • Boris: "Memory took a long time to crack... it feels like we finally got it right."
  • Product org: roughly 65% of PRs are now written by Tag, and climbing.
  • Cat: "You can just tell it, hey, remember to always do this thing, and it just has your back."

Conclusion

The through-line of the conversation is a shift from reactive, single-player AI to proactive, multiplayer AI. Long-horizon research means Claude can stay on track for days (and self-schedule beyond that), memory means it accumulates team-specific instructions instead of needing constant re-prompting, and living in public channels means best practices spread by observation. Inside Anthropic that combination has already pushed Tag to writing about 65% of the product org's PRs, and the roadmap — more platforms like Microsoft Teams, deep customizability — aims to put Claude within arm's length of every knowledge worker.

YouTube