Picture this: you're walking across a parking lot, running late for a meeting. You pull out your phone and type one line to Claude: "Pull together the Q1 network audit findings and draft the summary slide." By the time you sit down at your desk, the slide deck is waiting. Not because someone on the team stayed late. Because an AI agent on your Mac read the local files, opened the application, and assembled the output autonomously. Six months ago, that scenario was hypothetical. Today, every piece of the technology to make it work exists. That's not a product roadmap. That's a structural shift in how work initiates, executes, and completes.
That moment required four capabilities working together: persistent project memory, scheduled execution, phone-to-desktop dispatch, and visual computer control. In the span of four weeks, Anthropic shipped all four of these into Cowork. Scheduled Tasks launched on February 25, 2026. Dispatch arrived on March 17. Projects went live on March 20. And native Mac control for Computer Use landed on March 23. Four capabilities, one month, all composing into something greater than the sum.
That composition is the pattern worth paying attention to. These are four primitives of the same architectural layer, and together they answer a question we've been circling at INS since December: what does it look like when AI stops being a conversation partner and starts being operational infrastructure? We've been building toward this across our work on the agentic operating system, the new default, and the great convergence from doing to directing. Now we can point to the specific primitives that make it concrete.
Projects: Persistent Context Changes the Economics of Knowledge
Projects launched in Cowork on March 20, 2026. It's Anthropic's implementation of persistent, structured workspaces. Each project maintains its own knowledge base, custom instructions, and conversation history across sessions. You upload documents, define how Claude should behave within that context, and every subsequent conversation inherits that foundation automatically.
The mechanics matter here. Individual files can be up to 30MB across formats including PDF, DOCX, CSV, and HTML. Each project holds up to 200,000 tokens of context natively, and when that limit approaches, Claude activates retrieval-augmented generation to expand effective capacity by up to 10x. Team and Enterprise users can share projects with granular permissions: "Can use" for read-and-chat access, "Can edit" for full modification rights.
What this means practically: you build a project once for a client engagement, a product line, a compliance domain, or an internal process. Every conversation within that project starts with full context. No re-explaining. No re-uploading. No context rot from session boundaries resetting what the model knows about your work. The friction we used to accept as inevitable, spending the first ten minutes of every AI session re-establishing context, is no longer a fixed constraint. That time goes back to the actual work.
Projects turns AI from a stateless tool you re-brief every session into a persistent collaborator that accumulates institutional knowledge. The constraint shifts from "how much fits in one conversation" to "how well have we structured what the AI needs to know about this domain." That's a fundamentally different design problem, and it favors the people with the deepest domain expertise.
The architectural signal is clear. Anthropic is building toward AI that doesn't just answer questions but maintains working knowledge of your organization's context, terminology, constraints, and preferences. At INS, we see this as the natural extension of what we've been doing with Claude Code project structures for months. The people who invest in structuring that knowledge well will compound their advantage with every interaction. This is domain knowledge paying dividends in a new medium, and it rewards curiosity and organizational discipline over raw technical skill.
Scheduler: AI That Acts on Its Own Clock
Scheduled Tasks launched in Cowork on February 25, 2026, and the capability now exists in two forms. Both matter. In Cowork, you type /schedule inside any task and define a cadence: daily, weekly, weekdays, hourly, or on demand. You give it a name, a prompt, a working folder, and a model preference. It runs automatically. In Claude Code, the /loop command accepts standard five-field cron expressions and supports up to 50 concurrent scheduled tasks per session.
The execution model is deliberate. The scheduler checks every second for due tasks but enqueues them at low priority, firing between your turns rather than interrupting mid-response. Recurring tasks auto-expire after three days to prevent unbounded operation. Cloud-based scheduled tasks run on Anthropic's infrastructure, meaning they persist even when your computer is off.
Early adopters are already running zero-human-input workflows: daily engineering briefs that summarize overnight commits, nightly code audits that flag security patterns, automated dependency update checks that surface breaking changes before they reach production. These aren't theoretical use cases. They're patterns people have been running for weeks. And the people running them aren't losing work. They're gaining hours of judgment-intensive time that was previously consumed by routine synthesis. That's not less human involvement. It's better-allocated human involvement.
Scheduler is the difference between AI you use and AI that works. A tool you open when you remember to ask a question is useful. A system that monitors, summarizes, audits, and reports on a cadence you define is infrastructure. The distinction is not semantic. It changes staffing models, response times, and what "coverage" means for a team of any size.
For organizations already thinking in terms of agent primitives, Scheduler is the temporal primitive. It answers the question: when does this work happen? And the answer is no longer "when someone remembers to do it."
Dispatch: The Phone Becomes a Remote Control for Your Desktop Agent
Dispatch launched on March 17, 2026 as a research preview inside Claude Cowork. The concept is direct: you send a task from your phone, and Claude executes it on your desktop computer. Setup takes under two minutes. Download Claude Desktop, open Cowork, click Dispatch, scan a QR code with your phone.
The phone acts as a walkie-talkie to your desktop Cowork session. All processing happens locally on your computer. Claude accesses local files, connected applications, and installed plugins to complete whatever you've asked for. If it has the right integration for a task, it uses the API. If it doesn't, it falls back to Computer Use and navigates the application visually, the same way you would.
Dispatch is currently available to Max subscribers at $100 per month, with Pro access at $20 per month rolling out within days. It requires a Mac, and the computer must stay awake with Claude Desktop running. These are real constraints today. They also describe a research preview, not a finished product. If we've learned anything from watching this space, it's that today's constraint list is next quarter's changelog.
Dispatch redefines what "being at your desk" means. The person directing the work and the machine executing the work no longer need to be in the same room, or even on the same device. You carry the directing capacity in your pocket. Your desktop carries the execution capacity. This is not remote work. This is remote agency, the ability to initiate and oversee meaningful work output from anywhere, at any time.
The implications compound when you combine Dispatch with Scheduler. You can set up recurring tasks from your phone, monitor their outputs when convenient, and intervene only when your judgment is needed. The 10-80-10 model we've discussed before, 10% human ideation, 80% AI execution, 10% human taste and integration, now operates across devices and across time zones.
Computer Use: The Universal Integration Layer
Computer Use has been evolving since Anthropic first previewed it with Claude 3.5 Sonnet in October 2024, and the March 23, 2026 release brought native Mac control to Pro and Max subscribers. The trajectory tells the story more clearly than any description. On OSWorld, the benchmark that tests AI models on real-world computer tasks, Anthropic's scores have progressed from 14.9% to 28.0% to 42.2% to 61.4% to the current 72.5% with Sonnet 4.6. That's a five-fold improvement in eighteen months.
As of March 23, 2026, Claude can control your Mac directly. It moves the mouse, types, scrolls, captures screenshots, and navigates GUI applications autonomously. When Claude has an API integration for a task, it uses that. When it doesn't, it falls back to visual screen control, operating the application the way a human would. This fallback pattern is the important part.
Every enterprise runs dozens of applications that will never have AI-native APIs. Legacy systems, specialized vertical software, government portals, vendor dashboards. We've all accepted this as a permanent constraint, a category of work that simply couldn't be automated because no one was going to build an API for a vendor portal used by fifty people. Computer Use dissolves that constraint. It's not RPA, not brittle scripts that break when a button moves two pixels. It's a model that understands what it's looking at, reasons about what it needs to do, and adapts when the interface changes. The difference between scripted automation and visual reasoning is the difference between a macro and a colleague. And it means every application your team touches today is an application your agents can operate tomorrow.
Computer Use eliminates the integration tax. Every application your team uses is now an application your AI agent can use. Not through months of API development. Not through middleware. Through the same interface your people already use: the screen. This doesn't make API integrations unnecessary. It makes them optional, and it makes every application accessible to agent-directed workflows starting now.
The Architectural Pattern: Four Primitives, One Operating Layer
Zoom out from the individual features and the pattern becomes unmistakable. Anthropic is assembling the four foundational primitives of an agent operating layer:
Projects
Persistent context that accumulates and compounds across sessions. The agent remembers what it knows about your work.
Scheduler
Autonomous temporal execution. The agent acts on its own clock, not just when prompted.
Dispatch
Cross-device command and control. The agent takes direction from wherever you are.
Computer Use
Universal application access. The agent operates any software through visual reasoning.
Memory, time, direction, action. These are not features. They are the four prerequisites for AI that operates as infrastructure rather than software. An agent that remembers context, acts on schedule, takes remote direction, and can operate any application is structurally different from a chatbot. It's the difference between a tool you pick up and a system you configure.
We wrote in our piece on the agentic operating system that the operating system that matters most is no longer the one that manages your files but the one that manages your agents. These four primitives are Anthropic's answer to what that agent-managing layer actually looks like in practice.
The Monday Morning Questions: A Readiness Framework
Understanding the features is step one. Knowing where to apply them in your own organization is where the value compounds. We've distilled this into five questions you can bring to your team this week. Each one maps directly to one or more of the four primitives, and each one opens up a concrete conversation about where to start.
What knowledge lives only in people's heads?
Projects rewards organizations that have externalized their institutional knowledge into documents, guides, and structured references. If your most critical process knowledge is tribal knowledge, Projects will underperform until you fix that. The good news: Claude itself is an excellent tool for pulling tacit knowledge into written form.
What recurring work gets done late or inconsistently?
Scheduler is most valuable for tasks that someone is supposed to do on a cadence but that slip because of competing priorities. Audit summaries, status rollups, dependency checks, documentation updates. If you can describe the task precisely enough to delegate it to a new hire, you can schedule it.
Where does physical presence still bottleneck decision-making?
Dispatch matters most where the gap between "knowing what needs to happen" and "being at the machine to make it happen" creates delays. Field engineers, traveling executives, remote teams managing on-site infrastructure.
Which applications have no API and never will?
Computer Use is the answer for every vendor portal, legacy system, and specialized tool that your team accesses through a browser or desktop application with no programmatic interface. List them. That list is your Computer Use opportunity map.
What could your best people do if routine work disappeared?
This is the abundance question, and it's the most important one on this list. Every hour freed by scheduled tasks, persistent context, and automated application workflows is an hour your most experienced people can spend on the judgment-intensive work that actually compounds: client relationships, strategic thinking, mentoring, innovation. The constraint shifts from capacity to imagination. And imagination, unlike time, expands the more you use it.
The Compounding Effect
Each of these features is useful in isolation. Together, they compound in ways that are worth thinking through carefully.
A Project that contains your network documentation, vendor specs, and compliance requirements can be referenced by a Scheduled task that runs nightly audits against your configuration management database. Findings from that audit can be surfaced through Dispatch to a field engineer's phone while they're on-site. And if the remediation requires interacting with a vendor portal that has no API, Computer Use handles the execution.
That's not a hypothetical workflow. Every component of it works today. The compounding effect is the same pattern we've been documenting since our earliest posts on domain knowledge compounding: each primitive you integrate makes every other primitive more valuable. Projects make Scheduler smarter because scheduled tasks run with full context. Dispatch makes Computer Use more accessible because you can direct visual automation from anywhere. Scheduler makes Projects more valuable because persistent context gets refreshed and maintained automatically.
The organizations that will gain the most from these features are not the ones that adopt them fastest. They are the ones that combine them most thoughtfully. A scheduled task without good project context is just a recurring prompt. A Computer Use workflow without persistent memory re-learns your environment every session. The compounding happens at the intersections.
What This Means for INS
What This Means for INS
Internal impact: At INS, we already run our development workflows through Claude Code, and this blog is produced using an agent-assisted content pipeline with structured skills and iterative refinement. Projects maps directly onto how we organize client engagements and internal tooling. Scheduler is already part of our Claude Code workflow through the /loop command. These features don't change what we do. They formalize the infrastructure patterns we've been building toward.
The real acceleration is in Dispatch and Computer Use. Our team operates across multiple sites, and the ability to direct desktop agent workflows from a phone while on-site at a customer facility changes the response calculus. Configuration lookups, documentation pulls, vendor portal interactions, all of these become directable from the field without requiring a laptop or VPN tunnel.
Customer impact: For our OT and industrial networking clients, the implications are structural. Industrial environments run on specialized software that will never have modern APIs. SCADA interfaces, PLC programming tools, network management consoles, vendor-specific configuration utilities. Computer Use turns every one of these into an agent-accessible application. Scheduled agents can run nightly compliance checks against network configurations. Persistent project context can maintain the full history of a site's network architecture, change logs, and vendor specifications across every support interaction.
This is not about replacing the network engineers and OT specialists who understand these environments. It is about amplifying their reach. An engineer with deep knowledge of industrial protocols and site-specific configurations, combined with an agent that can execute routine checks, pull documentation, and navigate vendor portals autonomously, creates capacity that didn't exist before. The domain expertise becomes worth more, not less, because it can now be applied at a scale that was previously impossible for a single practitioner.
The Slide Deck That Built Itself
Come back to the parking lot scenario. One sentence from a phone. A slide deck assembled on a desktop from local files accumulated over weeks. That interaction requires all four primitives working together: persistent project context that knows which files matter, a dispatch channel from phone to desktop, computer use capability to open the application and build the slides, and the implicit scheduling of "do this now and have it ready by the time we sit down."
Six months ago, that workflow was science fiction. Today it works, with constraints and rough edges, but it works. The trajectory from 14.9% to 72.5% on computer use benchmarks tells you where the rough edges are heading. And here's the encouraging part: the people who start structuring their knowledge now, who begin externalizing institutional expertise into project contexts, who identify the recurring tasks worth scheduling, who map the applications that need visual automation, those people compound their head start with every improvement that ships. Every project you build today gets smarter with tomorrow's model. Every workflow you make agent-legible today becomes more capable as computer use scores climb from 72% toward 90%.
We are all learning together. And without exception, the organizations that lean in with genuine curiosity find themselves farther along than they expected. The tools are here. The primitives are defined. The only remaining variable is imagination, and that has always been the one resource that expands the more you use it.
The Infrastructure Is Here
AI with memory, time, direction, and action is no longer a roadmap item. It is a research preview you can use today. The organizations that treat these four primitives as building blocks, not novelties, will define what operational capacity looks like in the next era of knowledge work. The constraint is no longer the technology. It is the imagination to apply it.