Two Platforms, One Paradigm
In Lesson 3, you established that knowledge workers -- not developers -- are the central figures in the enterprise agentic transition. Now the question becomes practical: which platform do you actually use to deploy a domain agent?
The enterprise agentic landscape in 2026 is shaped by two dominant platforms. They share the same underlying paradigm -- large language models executing structured tasks against enterprise data -- but they differ fundamentally in who they target, how they deploy, and what they assume about your organisation. Understanding both, their architectures, their strengths, and their limitations, is a prerequisite for every deployment decision this book will ask you to make.
This lesson introduces both platforms and gives you a decision framework to choose between them. By the end, you will not be guessing. You will be qualifying.
Anthropic Cowork: The Knowledge Worker's Platform
Cowork is a desktop productivity tool built for knowledge workers. Out of the box, Claude is a generalist. Plugins turn it into a specialist for your role, team, and company.
A plugin bundles the domain expertise, tool connections, and workflows a specific role needs. Anthropic and the community have built plugins for sales, legal, finance, marketing, data, customer support, and more. You install one in seconds. The entire plugin is file-based -- markdown and JSON, no code, no infrastructure, no build steps.
What a Plugin Does
| What's Inside | What It Does |
|---|---|
| Skills (SKILL.md) | Domain expertise Claude draws on automatically -- persona, principles, and constraints |
| Commands | Explicit actions you trigger (e.g., /finance:reconciliation, /sales:call-prep) |
| Connectors | Wire Claude to the external tools your role depends on -- CRM, email, project tools, data warehouses -- via MCP |
The SKILL.md is the part only you can write. It is a natural-language instruction document that tells the agent who it is, what it knows, what it must never do, and how it should respond. It is not code. It is not configuration in the traditional sense. It is your institutional knowledge, made executable.
Plugins arrive as ready-made packages. Your contribution is customising the skills -- encoding how your organisation actually works. Chapter 15 covers plugin anatomy in detail.
Production Connectors
Cowork ships with production-ready MCP connectors for the systems knowledge workers already use: Google Workspace (Drive, Gmail, Calendar), Salesforce, HubSpot, DocuSign, Slack, Notion, Jira, WordPress, Apollo, Clay, Outreach, and Similarweb. Industry-specific connectors serve financial data (FactSet, MSCI), legal workflows (LegalZoom), and other verticals. When you write a SKILL.md that says "pull the latest pipeline data from Salesforce," the connector handles authentication, API calls, and data formatting. You never see it.
The connector ecosystem is Anthropic's primary infrastructure investment. The open-source MCP standard means third-party developers can build connectors for any system, and the growing marketplace through which connectors are published and distributed is creating the same kind of developer community dynamics that have driven the expansion of other platform ecosystems.
Deployment Model
Cowork follows a product-led growth model. Adoption is bottom-up: one team tries it, sees results, tells the next team. Procurement happens at the team level -- a department budget line, not a board-level capital expenditure. This means you can start deploying agents within weeks, not months.
The Typical Cowork Adopter
The organisations that have adopted Cowork most deeply typically started with a single team, a single vertical, and a single agent -- and expanded from there. The profile is the technically sophisticated team inside a larger organisation: the FP&A team that deployed a financial research agent before the CFO had approved a formal AI strategy. The in-house legal team that stood up a contract triage tool six months before IT had finished its vendor assessment. The architecture practice that deployed a BIM coordination assistant because the project manager could not wait for a platform decision. These teams share a characteristic: they have domain expertise that they understand is deployable in agent form, and they are not willing to wait for institutional processes to catch up.
OpenAI Frontier: The Enterprise Transformation Platform
Frontier takes a fundamentally different approach. Where Cowork gives individual teams a powerful plugin, Frontier provides the entire organisation with a semantic layer -- a unified intelligence infrastructure that sits across all systems.
What a Semantic Layer Means
Think of a semantic layer as a translation service that understands the meaning of data across every system in the organisation. When the finance team says "revenue," the sales team says "bookings," and the operations team says "throughput," a semantic layer knows these are related concepts and can reason across them.
In practice, this means an agent can move from a customer complaint logged in the CRM to a refund authorisation in the ERP to a follow-up communication drafted in the email system -- without a human touching the handoffs between departments. The agent carries context across each boundary: it knows why the refund was triggered, what the customer's history looks like, and what the finance team's approval threshold is. That cross-departmental continuity is what a single-team agent cannot replicate, and it is the capability that justifies Frontier's enterprise-wide deployment model.
Frontier manages agent identity, permissions, and cross-department context centrally. An agent deployed through Frontier does not belong to one team. It belongs to the organisation and can access data and workflows across departments -- subject to governance rules defined at the enterprise level.
Deployment Model
Frontier is sold top-down through Forward Deployed Engineers and consulting firm partnerships (McKinsey, BCG, Accenture, Capgemini). Early adopters include HP, Intuit, Oracle, State Farm, Thermo Fisher, and Uber.
Procurement is enterprise-wide: capital expenditure, legal review, security assessment, executive sponsorship. The deployment timeline is measured in quarters, not weeks. But the payoff is proportional: Frontier deployments touch every department, not just one team.
Side-by-Side Comparison
| Dimension | Anthropic Cowork | OpenAI Frontier |
|---|---|---|
| Target buyer | Team lead, department head | C-suite, CIO/CTO |
| Architecture | Plugins (skills, commands, connectors) | Unified semantic layer |
| Agent scope | Single team or function | Cross-department, enterprise-wide |
| Adoption path | Bottom-up (product-led growth) | Top-down (enterprise sales) |
| Procurement | Team budget | Capital expenditure + legal/security review |
| Time to first agent | Weeks | Quarters |
| Knowledge assumption | Concentrated in a specific team | Distributed across the organisation |
| Governance model | Team-level permissions | Centralised enterprise governance |
The Decision Framework
When choosing a platform, ask three questions:
Question 1: Organisational Scope
Is the problem contained within one team or function, or does it span the enterprise?
- If one team has the knowledge, owns the data, and will use the agent, Cowork fits. A compliance team building a regulatory review agent is a Cowork deployment.
- If the problem requires reasoning across departments -- say, connecting sales pipeline data to finance forecasts to operations capacity -- Frontier is designed for that cross-functional intelligence.
Question 2: Procurement Reality
Can you fund this from a team budget, or does it require capital expenditure?
- If your department can allocate budget without board approval, Cowork's product-led model lets you move fast. You deploy, measure value, and expand.
- If the investment requires a legal review, security assessment, and executive sponsorship, you are already in Frontier territory. The procurement process matches the platform's deployment model.
Question 3: Nature of Knowledge
Is the domain knowledge concentrated in one team or distributed across the organisation?
- If one expert or small team holds the knowledge needed to instruct the agent -- a senior architect who knows building codes, a compliance officer who knows regulatory requirements -- Cowork's SKILL.md model works. That expert writes the instructions.
- If the knowledge lives in multiple systems and no single person holds the complete picture, Frontier's semantic layer can synthesise across sources in ways that a single SKILL.md cannot.
Where Most Readers Start
For most Part 3 readers, Cowork is the starting point. You have domain expertise concentrated in your team. You have a team-level budget. You want results in weeks, not quarters. All hands-on exercises in this book use Cowork.
This does not mean Frontier is irrelevant to you. Understanding Frontier helps you recognise when your organisation is ready for enterprise-wide deployment -- and positions you to lead that conversation when the time comes. The platform decision is revisited in Chapter 26 for readers who are advising organisations on longer-term architecture choices.
Try With AI
Use these prompts in Anthropic Cowork or your preferred AI assistant to explore these concepts further.
Prompt 1: Personal Application
I work in [describe your role and team -- e.g., "a 12-person HR team
at a mid-size financial services company"]. We are considering deploying
an AI agent to help with [describe a specific task -- e.g., "screening
job applications against our competency framework"]. Based on the
Cowork vs Frontier decision framework, help me determine which platform
fits our situation. Ask me the three questions (organisational scope,
procurement reality, nature of knowledge) and then make a recommendation.
What you're learning: You are practising the decision framework against your own organisational context. The AI helps you think through each question honestly rather than guessing.
Prompt 2: Framework Analysis
Here are two fictional organisations. For each one, tell me whether
Cowork or Frontier is the better starting platform and explain why
using the three-question framework.
Organisation A: A 200-person law firm where the litigation team
(15 people) wants an agent to review discovery documents. The team
has its own budget and the senior partner has deep knowledge of the
firm's review standards.
Organisation B: A 5,000-person healthcare company where the CEO wants
to connect patient intake, clinical documentation, billing, and
compliance into a single AI-driven workflow. The project has
board-level sponsorship and a $2M budget.
What you're learning: You are testing whether you can apply the framework to unfamiliar scenarios. Getting the right answer for both organisations confirms you understand the framework, not just the vocabulary.
Prompt 3: Domain Research
Research the current state of enterprise AI platform adoption in
[YOUR INDUSTRY -- e.g., "architecture and construction," "financial
services," "healthcare"]. What platforms are organisations in my
industry actually using? Are they following a bottom-up (team-level)
or top-down (enterprise-wide) adoption pattern? What does that suggest
about whether Cowork or Frontier is gaining traction first?
What you're learning: You are connecting the abstract framework to real adoption patterns in your industry. This grounds your platform decision in evidence, not theory.