How Claude Tag brings an AI agent into a team’s Slack channels

Cat Wu asks Boris Cherny how day-to-day work with AI has changed. He recalls starting with code completion, where he decided what line to write and the tool helped finish it. Models progressed to functions, files and features; he says Claude Tag can now take on a whole experiment or data analysis.

The workflow also shifted from one person working line by line, to one person managing several Claude sessions, to an entire team interacting with an agent that takes more initiative, Cherny says.

Cherny describes Claude Tag as an agent added to a channel. Rather than requiring a new prompt for every step, he says, it can join a discussion, perform work that may take days or weeks, follow up and remember instructions. Everyone in the channel can see and participate in the work.

Asked what research enabled this, Wu says Anthropic has worked for years on longer-running autonomous models. She cites METR evaluations and says the latest models can work for 16 hours at a time. In Tag, she says, the agent can also schedule follow-ups for days or weeks later.

Cherny describes his own long-running experiment: a Tag session checks data each day, reports results and occasionally submits a bug fix. He says he has sessions lasting days or weeks, and possibly one that has run for a month.

Cherny says memory is another important capability. A team might instruct Tag to monitor only certain issue categories; he says it can remember that channel preference and adjust when another participant expands the scope. He says getting memory to work well took years of effort on Claude Code.

Wu asks how a team can avoid an agent interrupting every thread. Cherny says Claude is trained to judge when to participate and can be told to intervene more or less often, retaining that preference for future work.

Asked about usage patterns, Cherny says customers are experimenting with pull requests, production debugging and data analysis after connecting the relevant tools. He describes a Q&A channel where an agent can answer a question and mark it as handled, as well as a feedback channel where it works on fixes.

Wu says she often posts a data question in a shared channel and tags Claude for a first pass. After doing this repeatedly, she asked it to respond every time instead. The speakers see remembered standing instructions as a change from having to repeat the same prompt.

Wu says that because work happens in public channels, coworkers can watch experienced users and carry their approaches into new projects. She says Anthropic and its customers have seen practices spread as people observe one another.

Cherny says Claude Tag appeared in more of the company’s feedback and data channels after people saw early users working with it. He contrasts an ongoing objective, such as preparing code changes for bugs in a channel, with repeatedly opening an AI tool and issuing individual requests.

Cherny and Wu say shared sessions let several people guide the agent rather than copying an individual response to colleagues. Wu says the team can give feedback, learn from other users and refine an output together.

Cherny says a channel interface can lower the barrier to contributing to code for people who are uncomfortable with terminals, Git and checkouts. Wu says public channels also let Tag monitor projects; as a product manager, she asks for daily status reports across several features.

Wu says an agent connected to authoritative company files can also help new employees find answers to routine legal or benefits questions. Rather than waiting for another team, she says, they can ask in the channel when they need the information.

Cherny says Tag has improved productivity internally and estimates that it writes about 65% of pull requests in Anthropic’s product organization, a share he says is rising. He attributes some of the value to asking the agent to work and moving on rather than checking every step.

Wu asks how Cherny chooses between Tag and Claude Code. He says he began with small fixes and simple data questions, then delegated more complicated work as he grew comfortable. He says Tag can verify work in a remote sandbox and use preferences remembered for a particular channel.

As one verification example, Cherny says Tag may fix a bug and post a video to Slack. Wu asks what comes next; the speakers say Tag launched in Slack and they want to bring it to other collaboration platforms, including Microsoft Teams.

Wu says the aim is to bring Claude close to where knowledge workers already collaborate. Cherny emphasizes that the product is customizable and says he is interested to see how organizations adapt it for their teams.