The newest wave of AI agents from OpenAI, Anthropic, Microsoft, Perplexity, and Amazon all target the same user: someone who wants coding power without touching the terminal. But beneath their similar surfaces lies a fundamental disagreement about who owns the agent's brain, where its memory lives, and what you can take with you when you leave. Understanding these differences matters far more than the marketing pitch.

On July 9, OpenAI announced ChatGPT Work, which runs on the new GPT-5.6 model and can open local files, edit Google Workspace and Microsoft 365 documents, and carry multi-step tasks through to completion. Anthropic released Claude Cowork around the same time, positioning it directly against ChatGPT Work. Both target non-coders who want agent capabilities without the terminal. Yet the two products differ in a way that most users never see: where the agent's state lives, who holds the credentials, and how much control you retain.

What Are the Four Control Planes That Define AI Agent Products?

Behind every AI agent product, four critical decisions shape how much freedom you actually have. These are not marketing features; they are architectural choices that determine lock-in. The four planes are where execution runs, where state is persisted, how authority is delegated, and where policy is enforced.

Anthropic describes its design philosophy as "decoupling the brain from the hands." The model that reasons runs separately from the sandbox where code executes. A session, an append-only log of every model call, tool call, and result, connects the two. Because the sandbox is kept separate from the reasoning engine, the agent can start thinking before any container exists, and the code it runs stays far from the developer's credentials.

Anthropic

The same four planes appear across all major vendors. AWS's AgentCore Runtime gives each session a dedicated microVM with isolated CPU, memory, and filesystem, metering compute usage precisely. Google routes governed traffic through its Agent Gateway, where Model Armor policies inspect configured ingress and egress flows. Microsoft assigns each hosted agent a dedicated Entra Agent ID and runs it in a per-session sandbox whose filesystem survives idle periods.

How Do Agent Archetypes Differ by User Type?

The industry has settled on four main user personas, each with different expectations about control. These archetypes reveal how vendors position their products and, more importantly, what they are willing to let customers control.

• Knowledge Workers: ChatGPT Work, Claude Cowork, Copilot Cowork, Perplexity Computer, and Amazon Quick all serve this group. The vendor operates the runtime and sells the agent as a delegation tool for someone who lives in documents rather than code. The user grants access and supervises results, but the vendor manages runtime and persisted state in most cases, except for Perplexity's local option.
• Power Users Who Self-Host: OpenClaw and Hermes are the reference examples, both open source and running on the operator's own machine. Self-hosting involves managing the control plane rather than full local custody, since both still allow access to a hosted model and storage of credentials for external services.
• Developers: Claude Code, OpenAI Codex, GitHub Copilot in agent mode, and the open-source OpenCode live here. Coding-agent execution extends from the local IDE to vendor-managed sandboxes and asynchronous cloud workers, making this archetype the most challenging to categorize clearly.
• Enterprise Workflows: Organizations run open agent frameworks such as LangGraph or CrewAI on managed, governed runtimes. ADK on the Gemini Enterprise Agent Platform, Strands on the Bedrock AgentCore, Microsoft Agent Framework on Foundry Agent Service, and Claude Managed Agents belong here. The vendor operates infrastructure; the customer configures identity, policy, and retention.

The persona sells the product, but the four planes decide the lock-in. A product can be packaged on the surface and programmable underneath. Copilot Cowork shows this clearly: it is a packaged knowledge-worker experience, yet it runs on a governed enterprise platform and inherits Microsoft's identity, compliance, and audit controls.

Steps to Evaluate Control and Portability in AI Agent Platforms

• Check Runtime Ownership: Ask whether the vendor or your organization controls where the agent executes. Vendor-cloud runtimes offer convenience but reduce portability; self-hosted or customer-managed runtimes give you more control but require infrastructure investment.
• Understand State and Credential Storage: Determine where session history and API credentials are persisted. If the vendor holds them, you may face friction when switching platforms. If you control them, you retain more independence and can migrate more easily.
• Review Policy and Identity Controls: Examine what identity systems, access policies, and data retention rules you can configure. Enterprise platforms typically offer more granular control here, while packaged experiences often lock these decisions at the vendor level.
• Assess Data Portability: Ask explicitly what you can export and take with you if you leave. Some platforms offer append-only logs and session exports; others make this difficult or impossible, creating vendor lock-in.

Deploying agents on a managed runtime is more like leasing a workshop than purchasing a tool. The landlord manages the building and supplies the power, but a long-term tenant installs their own locks, maintains their records, and takes their tools when the lease ends. What differentiates each offering is not the compute operator, since vendors handle nearly all of it. It is how much of those four planes a product leaves the customer to configure and export.

A packaged experience hands nearly all four planes to the vendor and returns supervision and a finished outcome. A programmable platform operates the infrastructure but lets the customer define identity, policy, and retention and move the code elsewhere. ChatGPT Work makes this point from the developer side, since OpenAI's new desktop app folds Chat, Work, and Codex into a single surface, though OpenAI has not detailed how far the runtime or credential store are shared beneath it.

The key insight is that the persona and the control planes are separate questions. You can have a packaged knowledge-worker experience running on a governed enterprise platform, or a developer tool that looks simple but hides complex infrastructure choices. The marketing message tells you what the product does; the four planes tell you what you actually own.