🧠 AI · 2:43 min

Claude Cowork and chat merge: the agent moves into the cloud

Anthropic folded Cowork into every Claude conversation. The real change is not the interface: long tasks now run in a cloud sandbox.

A weekly report is due at noon. At 11:40 you ask Claude what moved in the pipeline, tell it to flag anything that slipped, and ask for five slides for the leadership meeting. Then you close the laptop and walk into another meeting.

That scene was awkward for most of 2026. Chat and Cowork were separate products: one for questions, one for longer work, each with its own context and its own place in the app. On 16 September 2026 Anthropic merged them for Pro, Max and Ultra plans. The unified experience reached Claude Enterprise on 23 October and Team on 12 November. On 16 November it became the default for existing Enterprise organizations that had not opted out.

The interface change is easy to describe. The operational change is bigger: when a conversation turns into work, the job leaves your machine and runs in a cloud environment that Anthropic manages.

One conversation, two execution modes

The conversation still starts as a chat. A simple question gets a quick answer. When the task needs to run code, read files or produce a document, Claude continues in a cloud session and returns the finished analysis, document or deck. Chats, projects, connectors, skills and plugins carry over to the new experience.

Because the work runs in the cloud, closing the laptop does not stop it. You can check progress from the web app or the phone and pick up the result later. That is the real promise of the merge: not a new button, but continuity between asking and delegating.

Anthropic says each cloud session runs in its own isolated virtual machine. For Claude Enterprise, network egress is off by default and connector credentials are kept outside the sandbox. Those are sensible defaults, but they describe the platform, not your configuration. The connectors an admin enables and the roles that can use them still define what the agent can reach.

The bill is part of the architecture

Agentic work can use more tokens than a quick chat. A task that reads a dozen files, checks a spreadsheet and drafts a report consumes far more context than a question about those files. Anthropic's answer is a set of administrative controls rather than a single price change: an Efficient mode that keeps the unified experience on Sonnet and Haiku models, model entitlements per role, spend limits at organization, group and member level, and pooled group budgets.

Before a pilot, estimate the floor. This script uses an illustrative price; replace it with the number in your contract:

python
# Estimate the monthly token bill of a cloud-agent pilot.
runs_per_day = 40
tokens_per_run = 180_000
price_per_million = 3.00

monthly = runs_per_day * 30 * tokens_per_run / 1_000_000 * price_per_million
print(f"estimated agent spend: ${monthly:,.0f} per month")

The output is a floor, not a forecast. Retries, long sessions and larger documents add up. The point is to put a number next to the pilot before the first invoice arrives.

The trust boundary moved

On a laptop, an agent could read what the local user could read. In a cloud session, the blast radius is defined by connectors, roles, retention settings and egress policy. That can be an improvement: controls are centralized and admins get visibility across the organization. It also means a task can touch systems that the person who started it never opened.

Prompt injection is the obvious risk. A document, an email or a web page can contain instructions that the model treats as part of the task. OWASP recommends least privilege, validation of tool inputs and human approval for high-impact actions. A unified interface does not remove that work; it makes the approval step more important, because the person delegating the task may not see every action taken in the cloud.

Claude asks before taking an action by default, and users can choose to check in less often. Organizations should decide which choice fits each role. A finance team processing invoices needs different defaults from a design team drafting slides.

There are also configuration exceptions worth knowing. Organizations using zero data retention for Claude Code, a HIPAA-ready configuration or customer-managed encryption keys stay on their current setup. If that is you, the merge is not a reason to change a compliance control by accident.

What to check before turning it on

Anthropic published an admin setup guide alongside the Enterprise rollout, and the Help Center lists the capabilities included in the unified experience. Read both before a broad launch.

The open question

The merge removes a decision users found annoying: which Claude should I use? It replaces that decision with a quieter one: do I know when my work is running in the cloud, on which systems, at what cost? Teams that answer those questions before the rollout will get the convenience without losing the boundary.

Sources