A Channel Is Not an Assignment

Anthropic just put Claude in Slack. You tag @Claude in a channel, hand it a job in plain language, and it works alongside the team like another headcount. Claude Tag is pitched as a virtual employee: multiplayer, always on, learning your company one message at a time. Anthropic says 65 percent of its own product team’s code now comes from the internal version. The model is good enough. The demand is real.

There is an easy complaint here, that chat is a clumsy interface for an AI, and I am going to skip it because it has been made a hundred times. The narrower problem is the one that bites. The job being handed to Claude is work assignment, and work assignment is something we solved years ago with a specific tool that is not chat. Slack is the place we already know loses the work. Handing that job to an agent makes it worse.

Work Assignment Is Already Solved

Jira, Linear, GitHub Issues, Asana, ServiceNow, and a hundred others all exist to do the one thing chat and email do badly: assign a piece of work and track who owns it until it is done. An entire software category formed in that gap. We built it because chat kept losing the work. Filling out forms was never the fun part.

A ticket is a small durable contract. One owner. A discrete scope. A definition of done. A state you can read at a glance: open, in progress, in review, closed. An audit trail attached to the unit of work instead of smeared across a room full of conversations. You can ask a tracker “what is assigned, to whom, and what is blocked” and get a straight answer.

A chat message carries none of that. It has a sender, a timestamp, and a spot in a scroll. “@Claude can you take the export bug” lands between lunch plans and a deploy alert, collects two emoji, and rides up out of view. No owner of the outcome. No agreed definition of done. No state to check. The only record is a thread you reassemble by reading.

Humans Paper Over the Cracks. Agents Don’t.

Here is the part the tired version of this argument misses. Chat has always been a bad place to assign work, and it mostly gets away with it, because people quietly cover for everything it lacks.

Drop a vague task into a channel and a decent colleague reads the room. They notice nobody picked it up and raise a hand. They ask whether you want them to own it end to end. They sense when a thread has gone quiet and nudge it back to life. They tell you when it is done. The ticketing system only writes down, formally, what good teammates were already doing on instinct. Chat leans on that instinct the entire time and never admits it.

An agent brings none of it. No read on the room. No impulse to claim ownership. No nudge to confirm scope before it starts. No sense that a thread died three days ago with the work half-finished. Every social reflex that quietly rescues chat for humans is exactly the thing the model does not have. So the cracks people step over stay open holes. The agent needs the structure a tracker provides more than a person does, and chat hands it less of that structure than any tool you own.

Claude Tag then takes chat’s weakest properties and ships them as features. It is multiplayer: one Claude per channel, and anyone can pick up where the last person left off. For accountability that is the bystander effect rendered in software. When a task belongs to a channel it belongs to nobody, and now the only worker guaranteed never to grab it is the one you assigned it to. Using AI is management, and the first rule of delegation is one owner with one definition of success. A channel gives you neither, and the agent in it will not supply what is missing.

The Best Prompt Was Always a Ticket

I have made the case that the best prompts look like acceptance criteria on a well-written ticket. Here is what done looks like. Here is what I will reject. Here are the constraints. Go. We worked out that the right unit of agent work is a ticket, then turned around and put it inside a chat box.

This is the English Trap wearing a new coat. Talking to a model is specification. Chat is the one surface built to make it feel like a conversation. Slack tells your hands and your instincts you are messaging a colleague, the single most expensive misread you can have with a language model. A ticket tells the truth. You are writing a spec, and something is going to be held to it.

Done Has to Be Checkable

We automate what we can verify. A ticket makes “is this done” cheap to answer: there are acceptance criteria, and a state that flips to closed when they are met. A thread makes the same question expensive. Doneness is a judgment buried somewhere in the replies, if it is anywhere at all. Multiply that across every task in the channel and checking status becomes an archaeology project.

Ambient mode pushes it the wrong way again. Switch it on and Claude interjects on its own, surfaces things it thinks you should see, chases threads that went quiet. An agent with no scope and no success condition is the run-forever shape that agents are supposed to avoid, now running loose in your team’s main room. The point of a tight task is that the model can tell when it is finished. Ambient proactivity is the standing instruction to never finish. Two years of learning not to let the model wander, reintroduced as ambience and aimed at you.

They Rebuilt the Tracker Inside Slack

Read the rest of the feature list and you can watch Anthropic hit the problem. Activity logging. Task-requester attribution. Per-channel spend limits. Scoped identities so the sales agent cannot read engineering’s data. Every one of those is something a tracker has on day one, rebuilt inside chat after the fact because the substrate will not supply it. When you find yourself reconstructing ownership, scope, and an audit trail on top of a tool, the tool was the wrong base. Correct by construction beats bolted on, and this is bolted on.

So why reach for chat at all? Distribution. Slack is where the users already sit, and tailing the conversation is how you accumulate the org’s context, which several observers have correctly named the real prize. Putting the agent in the channel is a customer-acquisition decision wearing a workflow decision’s clothes. Good strategy for the vendor. It does not make chat the right home for your work.

Put the Agent Where the Work Is

The integration worth wanting is Claude as an assignee in the tracker.

Assign it a ticket. It works the ticket, moves it to review with the output attached, and a human accepts or rejects against criteria written down before any work started. Rejected work goes back with a reason, and the loop closes against the same acceptance check, which is the only reason agents converge at all. The state machine already exists. The ownership model already exists. The audit trail already exists, sitting on the ticket where you will look for it later. Nothing to bolt on, and nothing left for a missing social instinct to cover.

None of this makes chat useless. For a quick question, a “what does this stack trace mean,” a half-formed idea you want to push on, chat is the right tool and always was. The mistake is promoting chat to system of record for owned, deliverable work. Conversation and accountability are different jobs, and Anthropic just shipped the first one wearing the second one’s badge.

The agent is a new kind of worker, and the most literal-minded one you have ever managed. It will not read the room, claim the task, or close the loop on its own. So give it the plumbing we built precisely so none of that has to be left to instinct: a place where a task has an owner, a definition of done, and a state that ends in closed. That place already exists in your company, and it is not the chat window.