Claude's 2026 Stack

    Five products. Five levels of unlock. Start where you are. Climb as you go.

    Claude 2026:
    Five Levels of Unlock

    Anthropic didn't just release new models. They shipped an end-to-end agentic operating layer. Each tier unlocks a different kind of work. This guide walks every level, shows you where you are today, and maps the path to the ones you haven't reached yet.

    L01

    Cowork

    The deskside agent

    L02

    Scheduled Tasks

    Recurring work

    L03

    Dispatch

    Phone as remote

    L04

    Routines

    Cloud-native

    L05

    Managed Agents

    Production agents

    Level 01

    Cowork: The First Door Open

    Claude Code for the rest of your work. A deskside agent that reaches into your files, apps, and connectors. Accessible to everyone with Claude Desktop.

    Anthropic's own framing

    "Claude Code for the rest of your work."

    Same agentic engine. Different wrapper. Cowork is task delegation in a GUI, sitting inside Claude Desktop. You hand over the goal. Claude figures out how to get it done.

    Local file access

    Read, create, and edit files on your machine within a folder you authorise.

    Connectors

    Gmail, Drive, Notion, Slack, Linear, and other services via MCP.

    Sub-agents in parallel

    Spawn specialists for complex work. Claude coordinates them for you.

    Plugin marketplace

    Domain plugins for legal, sales, finance, marketing, HR, design, and more.

    Projects

    Persistent workspaces with saved prompts, files, and prior context.

    Computer Use (preview)

    Claude can point, click, and operate your screen to finish the job.

    The key distinction

    Cowork delegates. Claude Code executes.

    Cowork is "here's the goal, figure it out." Claude Code is "run this command, take this action." Both sit on the same Claude Agent SDK. Different surface, same engine.

    Level 02

    Scheduled Tasks: Unlock Recurring Work

    Describe a job once. Cowork runs it on your cadence while you focus elsewhere. The first unlock that buys back your time.

    Pick a cadence. Cowork handles it.

    Example: daily - Cowork fires the task, uses connected tools, drops the result where you told it to.

    Daily briefings

    Every morning 7am

    Summarise Slack, emails, and calendar from the past 24 hours. On your desktop when you sit down.

    Weekly reports

    Monday 8am

    Compile data from Drive, spreadsheets, and connected tools into a formatted summary.

    Recurring research

    Daily at 6am

    Track competitors, topics, or industry news on a cadence you set.

    File organisation

    Weekly

    Sort, clean, or process files in a designated folder while you sleep.

    Team updates

    Weekdays 9am

    Generate status updates or standup summaries from your PM tools.

    The catch

    Scheduled Tasks only run while your desktop is awake and Cowork is open.

    Close the lid and the task sleeps with the machine. For always-on, cloud-native scheduling, skip ahead to Routines.

    Level 03

    Dispatch: Unlock Your Phone

    A persistent thread from mobile to desktop. Fire the instruction from anywhere. The result waits for you.

    Not screen-sharing. Not VNC. Your phone fires an instruction. Your desktop runs it. The result waits for you.

    The relay

    01

    You tap your phone

    "Summarise my top 10 unread emails and flag anything urgent."

    02

    Dispatch relays

    Persistent thread pushes the instruction to your desktop in real time.

    03

    Desktop executes

    Cowork spins up a session, uses connectors, plugins, and local files.

    04

    Result + notification

    Deliverable waits on your desktop. Push notification lets you know.

    Real-world patterns

    Trigger a research agent from bed
    Build a 15-slide deck while in a cab
    Clean your inbox on a flight
    Process CSVs at dinner
    Draft a client memo from the car
    Kick off a competitive scan before a meeting

    Same catch as Scheduled Tasks

    Desktop has to be awake and Cowork has to be open. Dispatch is a remote control, not cloud execution. If you need truly laptop-independent automation, that's Routines.

    Enjoying the guide?

    Enter your details to unlock the remaining insights. We'll also send you the complete guide as a reference.

    No spam. Unsubscribe anytime.

    Level 04

    Routines: Unlock The Cloud

    Claude Code configs running on Anthropic's infrastructure. Schedule, API, and GitHub triggers. Laptop-independent.

    Cloud-native. Laptop-independent.

    A Routine is a saved Claude Code config (prompt, repos, connectors) that runs on Anthropic's cloud. Close your laptop. The work keeps going.

    Cron expressions, without the ops headache

    Hourly, daily, weekdays, weekly from the UI. Custom intervals via the CLI. Anthropic runs the cloud.

    schedule trigger
    $ claude routines create \
        --name "docs-drift" \
        --schedule "0 2 * * 1" \
        --repo org/product-docs \
        --prompt "Scan merged PRs, flag
                 outdated docs, open PRs"
    

    Backlog maintenance

    Nightly. Labels issues, assigns owners, posts Slack digest.

    Docs drift

    Weekly. Scans merged PRs, opens updates for outdated docs.

    Level 05

    Managed Agents: Unlock Production

    A hosted agent runtime. Sandboxed containers, persistent sessions, SSE streaming. Anthropic runs the plumbing so you ship product.

    The pitch

    Stop building your own agent loop.

    Before Managed Agents, shipping a production agent meant writing the loop, managing tool execution, setting up containers, handling caching, context windows, retries, streaming, and networking. 80% infrastructure, 20% intelligence. Managed Agents gives you a hosted runtime so you only write the agent.

    Agent

    Reusable config

    Model, system prompt, tools, MCP servers. Creates an agent_id you can re-use across sessions.

    Environment

    Container template

    Packages (pip/npm/apt), networking rules, env vars, setup scripts, GitHub repo mounts.

    Session

    Running instance

    Each session is an isolated container with its own filesystem, processes, and network.

    Events

    SSE comms

    Send user messages, stream back responses, tool notifications, status updates. Interrupt mid-flight.

    Multi-agent (research preview)

    Coordinator delegates to specialist workers in named threads

    agents.py
    coordinator = client.beta.agents.create(
        name="Tech Lead",
        model="claude-opus-4-6",
        agents=[
            {"agent_id": reviewer.id,
             "name": "Code Reviewer"},
            {"agent_id": tester.id,
             "name": "Test Writer"},
        ],
    )

    Workers share the container and filesystem but operate in isolated conversation contexts. One level of delegation today: coordinator dispatches, workers don't re-dispatch.

    Infrastructure Anthropic handles

    • - Sandboxed execution with strict permissions
    • - Parallel vs sequential tool decisions
    • - Error handling, retries, idempotency
    • - Prompt caching and context compaction
    • - Background execution + session persistence
    • - Config versioning with rollback

    Pricing shape

    • - $0.08 per session-hour, metered to the millisecond
    • - Token pricing matches the standard API
    • - Cache reads at 0.1x input token price
    • - Web search: $10 per 1,000 searches
    • - 1-hour Opus session, 50K in / 15K out ~ $0.70

    Architecture

    The Local vs Cloud Split

    The single biggest decision in this stack. Which levels live on your machine, which live in Anthropic's cloud, and why the difference matters.

    The single most important architecture decision: does the work run on your machine or in the cloud?

    Local

    Runs on your machine

    Full access to local files, apps, and screen. Requires the desktop awake and Cowork open.

    Cowork- Desktop GUI agent
    Scheduled Tasks- Local cron, inside Cowork
    Dispatch- Remote control for your local session

    Close the lid. Work pauses.

    Cloud

    Runs on Anthropic's infrastructure

    Always-on, laptop-independent. Sandboxed containers, session persistence, built for production.

    Routines- Cloud-native scheduled and event-driven jobs
    Managed Agents API- Hosted runtime for product-facing agents

    Close the lid. Work continues.

    They compose. Most mature teams use both: Routines for reliable scheduling in the cloud, Managed Agents for deep product-facing reasoning, Cowork plus Dispatch for human-in-the-loop review from anywhere.

    Playbook

    Use Cases By Persona

    Same stack. Different jobs. Here's the unlock pattern for solo founders, developers, agencies, and enterprise ops.

    Same stack. Different audiences. Here's the unlock pattern for each persona.

    Solo founders

    You wear every hat. These unlocks give you output without adding headcount.

    Morning brief

    Scheduled Tasks

    Cowork fires at 7:30am. Reads Gmail, Slack, calendar, and priority files. Daily briefing waits on your desktop.

    Mobile research trigger

    Dispatch

    Fire a research agent from your phone mid-commute. The full competitive analysis is ready when you arrive.

    Newsletter automation

    Routines

    Routine fires weekly. Reads inbox for content ideas. Drafts newsletter structure. Posts to your draft queue.

    Content repurposing

    Cowork

    Cowork task takes a finished long-form piece and breaks it into 10 social assets, captions, and subject lines.

    Decision

    Pick Your Level

    Eight scenarios. One recommendation each. The right tier falls out of the job.

    Pick the scenario. The right tool falls out.

    Pick this

    Cowork

    GUI-driven, no code required. Access to local files, connectors, and plugins out of the box.

    Advanced

    The Power Stack

    Compose the layers: Routines triggers, Managed Agents reasons, n8n routes, Dispatch reviews.

    Mature teams don't pick one. They stack them. Here's the composable architecture.

    01

    Routines

    The scheduler

    Fires on a schedule or a GitHub/monitoring webhook. Laptop-independent.

    02

    Managed Agents

    The reasoning engine

    Does the heavy work. Called via API by the routine. Sessions run in sandboxed containers.

    03

    n8n / Make

    The integration layer

    Routes results to non-Anthropic systems. CRM, Slack, databases, billing. Compliance and governance wrappers.

    04

    Cowork + Dispatch

    The human review layer

    Outcome notification lands on your phone. Approve, steer, or override from anywhere.

    What this buys you

    • - Cloud-native automation with no laptop dependency
    • - Deep Claude reasoning via Managed Agents
    • - Multi-vendor integration and compliance controls from n8n
    • - Asynchronous human oversight via Dispatch on your phone
    • - No custom agent infrastructure to build or maintain

    Journey

    How We Got Here

    Twelve releases in four months. Most of these didn't exist in December. All of them exist now.

    Twelve releases. Four months. The shape of the 2026 stack, in the order it landed.

    Cowork launches on Max (macOS)

    Claude's general agentic assistant arrives inside Claude Desktop. Max subscribers on macOS get access first. The deskside agent era begins.

    Cowork expands to Pro

    Pro plan picks up Cowork within the first week. Broad SMB and consumer access opens up.

    Cowork on Team and Enterprise

    Organisations get Cowork with admin controls. Enterprise rollout begins.

    Opus 4.6 released

    New flagship model lands. Available across Claude products.

    Cowork on Windows

    Windows desktop support ships. The full Cowork experience arrives cross-platform.

    Sonnet 4.6 + 1M context beta

    Sonnet 4.6 and 1M-token context beta land together. Connectors open up on the free tier.

    Scheduled Tasks in Cowork

    Describe a recurring job once. Cowork runs it on your cadence. Hourly, daily, weekdays, weekly.

    Dispatch launches

    Persistent thread from Claude mobile to Claude Desktop. Your phone becomes a remote for your Cowork session.

    Computer Use in Cowork + Claude Code

    Claude can point, click, and operate your screen directly. Research preview.

    Managed Agents API (beta)

    Pre-built, configurable agent runtime on Anthropic's infrastructure. No more rolling your own agent loop.

    Cowork GA: analytics, OpenTelemetry, RBAC

    Cowork hits general availability with enterprise observability and access controls.

    Routines in Claude Code (research preview)

    Cloud-native scheduled and event-driven automation inside Claude Code. Laptop-independent. Anthropic runs the infrastructure.

    Most of these didn't exist in December. All of them exist now. The pace isn't slowing down.

    AnthropicAnthropichashasshippedshippedananagenticagenticoperatingoperatinglayer.layer.CoworkCoworkforforknowledgeknowledgeworkers.workers.RoutinesRoutinesandandManagedManagedAgentsAgentsforforbuilders.builders.DispatchDispatchforforwhenwhenyou'reyou'reawayawayfromfromthethedesk.desk.PickPickthethetier.tier.ShipShipthethesystem.system.

    Cowork

    Deskside agent for knowledge workers

    Scheduled Tasks

    Local recurring jobs inside Cowork

    Dispatch

    Phone-to-desktop delegation

    Routines

    Cloud-native dev automation

    Managed Agents

    Hosted runtime for product agents