SUMMARY
Claude Dispatch & Computer Use: Getting Real
Work Off Your Desk
Source: Video transcript  |  Speaker: Nate (AI productivity commentator)  |  Topic: Anthropic product
launches — Dispatch, Scheduled Tasks, Computer Use
The Core Framing: Work On vs. Off the Desk
Anthropic shipped two things in quick succession — Dispatch and Computer Use — and
together they constitute what OpenClaw users have had, but without the self-hosting
overhead. Claude can be texted from a phone, take over a desktop, open apps, click through
screens, navigate tools with no API, and deliver finished work while the user is elsewhere.
The output is not a summary to review or a draft to edit — it is the finished thing.
Most AI agent demos are optimized to look good, not to remove work from someone's plate.
The recurring failure mode is agents that produce proactive briefings — more documents to
read. The relevant test is whether the agent's output clears something that would otherwise
sit on the desk generating cognitive load.
Primitive 1: Scheduled Cloud Tasks
Scheduled tasks in Claude are not the same as scheduled tasks in chat. They run on
Anthropic's infrastructure — a controlled cloud environment with configurable network
access, environment variables, and setup scripts — whether or not the user's laptop is on.
The model is: a repository, a schedule, and a prompt. Anthropic runs the rest.

Anthropic uses this internally to keep a Go and Python library in sync — a codebase in
one language automatically mirroring another on a schedule. This replaces several hours
per week of important-but-never-urgent engineering work that typically falls through the
cracks.
Tasks can be scheduled every hour, two hours, or three hours. Real-time (per-minute)
monitoring is not the use case. The primitive connects to any MCP server already configured
in Claude — GitHub, Slack, Linear, Google Drive, OpenBrain — with no need to
reconfigure.
Non-developer examples: waking up to an AI news digest already processed and stored;
flight price alerts for a specific route when the price drops below a threshold; reminders for
bills that don't support autopay. Anything time-triggered with a clear output condition is a
candidate.
Primitive 2: Dispatch as an Orchestration Layer
Dispatch is broadly described as persistent chat for the phone — one long context window
that doesn't reset. That framing undersells the actual architecture. Pairing a phone with
Claude desktop via QR code makes the phone a command surface and the desktop an
execution surface. From one mobile conversation, multiple Claude co-work sessions can be
spawned and run simultaneously, each with its own context, file access, and connectors.
Work runs in parallel — and that is the point.
Product manager Pavle Hurin ran Dispatch for 48 hours straight. He spent roughly 25
minutes entering commands over two days. Claude ran competitor analysis, drafted
stakeholder messaging, and managed multiple rounds of iteration in parallel — while
Hurin watched his kids at a bounce house. His phone was the command interface; the
desktop executed.

The model this enables is management rather than supervision. A manager does not watch
over a worker's shoulder; they set clear direction, then step away. Dispatch makes that
pattern available for Claude co-work: spin up tasks, define intent clearly, go about the day,
check results when ready.
Current limitations: each subtask requests folder access individually (no bulk approval yet);
files cannot be attached from the phone or received back directly (workaround: sync co-work
to Google Drive or Dropbox); complex multi-app tasks succeed roughly half the time in
early testing; the desktop must stay on. These are research preview constraints expected to
narrow over time.
Primitive 3: Computer Use for Apps Without MCP
MCP coverage will never be complete. More than half the accessible web — legacy ERP
systems, old Jira instances, bespoke SAP screens — will never have clean APIs or
connectors. Computer use addresses this: Claude can operate the keyboard and mouse
remotely via co-work, navigating any app a human could navigate on screen.
The practical scenario this unlocks: manual data extraction from a legacy system that has no
API and can't get onto an engineering backlog, moved into a spreadsheet and sent to the
accountant — entirely delegated via Dispatch, executed via computer use, deposited in a
shared folder. The user never sits through the manual process.
MCP is the preferred interface — Anthropic encourages it, developers adopt it, and it
is the right first choice when available. But computer use removes the constraint that
only MCP-connected apps are within reach. The agent's accessible surface is now
essentially any software that runs on the desktop.

OpenClaw vs. Anthropic: Self-Hosted vs. Managed
The distinction is not primarily about safety — it is about who manages the infrastructure.
OpenClaw requires the user to configure servers, manage credentials, vet skill marketplaces,
and troubleshoot connections. Developers who want that control get significant raw freedom,
including the ability to swap in any LLM. For everyone else, it is a considerable
maintenance burden.
Anthropic's stack handles servers, sandboxing, permissions, and network configuration on
behalf of the user. The trade-off is lock-in: cloud-only, Claude-only, no local Ollama. But for
the majority of people who want capable agents without infrastructure work, the abstraction
is the product.
This follows a well-worn pattern: sendmail → Gmail; rack servers → AWS; Jenkins
→ GitHub Actions. Self-hosted versions prove the category exists; managed versions
achieve mass adoption. OpenClaw proved people want always-on persistent agents.
Anthropic is now shipping the version that scales.
A Framework for Delegating to Agents
Close open commitment loops. Every promise made to a client, team, or stakeholder that
hasn't been delivered is an open loop generating mental overhead. Revised scopes, meeting
minutes, memoranda — these are regular outputs that agents can produce with good
prompting and context. Quality failures are usually prompting failures, and prompting is a
learnable skill.
Improve decisions under time pressure. Walking into a meeting having seen 30% of
available information is common. Agents can be dispatched in the morning to pull the
relevant dashboard, summarize background docs, and surface data that wouldn't ordinarily

be consulted. The right use is expanding the information base, not asking AI to confirm a
pre-formed opinion.
Build compound signal detection. An agent with access to a persistent knowledge base
(e.g., OpenBrain via MCP) can surface patterns over time — connecting a competitor's
hiring spike to a strategy conversation from three weeks ago, or linking patent filings to
emerging market moves. This transitions the agent from reactive tool to proactive
collaborator.
Overnight engineering work. Scheduled cloud tasks make it practical to run long-horizon
engineering jobs without occupying a machine or a person: migrating dependencies,
improving test coverage to 80%, refactoring authentication layers — work that is always
important and never urgent enough to schedule during the day.
The Trust Shift: Learning to Walk Away
The practical bottleneck for agentic workflows in 2026 is not tooling — it is the human
instinct to hover. The impulse to go back and verify that the agent is actually working runs
against the entire value proposition of asynchronous delegation. Agents running in parallel,
executing without supervision, are more valuable the more confidently their operator can
step away.
The people who benefit most from these tools will be those who can genuinely disengage —
who set intent clearly, dispatch the task, and return to results rather than monitoring process.
That behavioral shift, not any feature release, is the real unlock of the second half of 2026.