Since ChatGPT arrived in late 2022, three patterns of human-AI interaction have become visible: Chat, Code and Claw. They aren’t clean historical eras. All three coexist, often inside the same product. What changes between them is who carries the work forward and how much authority the software receives.
I’ve watched these patterns converge while helping enterprises adopt Microsoft’s AI stack at Wortell. The names are deliberately simple. They make it easier to see why a chat rollout, a coding agent and a persistent personal agent require different controls.
Chat: “Help me write this”
The first mainstream pattern was conversational. ChatGPT launched in late 2022 as a dialogue system: you could ask follow-up questions, challenge an answer and keep context within a conversation. Draft an email. Summarise a document. Brainstorm marketing copy. The user still drove the exchange.
Microsoft embedded this paradigm across its entire productivity stack through Copilot. Wave 1 of Microsoft 365 Copilot brought chat-based AI into Word, Excel, PowerPoint, and Outlook. You could ask Copilot to summarize a Teams meeting, draft a report based on some emails, or generate a formula in Excel. It was genuinely useful, and for millions of knowledge workers, it was their first real encounter with generative AI.
Chat’s structural limitation wasn’t that every interaction had to be a single turn. It was that the user remained the orchestrator. You decided what to ask next, moved information between systems and judged when the task was finished. The model could hold a conversation, but it did not independently carry a job through your environment.
Chat proved that useful output could arrive at conversational speed. It also exposed a ceiling: a good answer is still only an answer. The user has to turn it into action.
Code: “Build it for me”
The second age arrived when AI learned to act, not just answer. In developer circles, this shift happened fast. GitHub Copilot evolved from an autocomplete tool into a full coding agent, one that can read your repository, understand your architecture, create a branch, write the code, run the tests, and open a pull request. You assign it an issue before lunch; by the time you’re back, there’s a PR waiting.
This is a different interaction pattern. Prompting is still involved, but the unit of work has changed from an answer to a delegated outcome. The agent makes a plan, uses tools, evaluates results and iterates.
In the Microsoft ecosystem, this pattern appears across several layers. Microsoft Foundry became generally available in March 2026 as a platform for production agents, including multi-agent workflows, persistent state and hosted tools. Foundry Agent Service ships SDKs for Python, JavaScript, Java and .NET.
On the low-code side, Copilot Studio has moved beyond chatbot building. Business users can design automation with natural language, while developers can add orchestration, connectors, MCP tools and human approval steps. Some agents answer interactively; others run when an event or schedule triggers them.
The practical change is access. Professional developers can build in Foundry and the Microsoft Agent Framework, while business users work in Copilot Studio. Both can now create software that takes actions instead of only producing text.
Claw: “Just handle it”
And then came the lobster.
In early 2026, Peter Steinberger’s open-source OpenClaw project drew attention to a more personal form of agent. It runs on hardware you control, connects to messaging channels such as WhatsApp, Telegram, Slack and Teams, and can use tools on your behalf.
OpenClaw can act from a message, a schedule or another configured trigger. People use it for calendars, code deployments, notes and smart-home tasks. NVIDIA’s NemoClaw work showed that the same operating pattern had already caught the attention of a major platform vendor.
Claw is the persistent version of the pattern. The agent stays connected to channels and tools, keeps state and can run from configured triggers. That is more precise than saying it somehow learns what you need before you ask; proactivity comes from the goals, schedules, events and permissions we give it.
Microsoft clearly sees this coming. Wave 3 of Microsoft 365 Copilot, announced in March 2026, introduces Copilot Cowork, built in partnership with Anthropic using Claude’s technology. Cowork can execute multi-step tasks that unfold over minutes or hours, reasoning across Outlook, Teams, Excel, and your files. It breaks down complex requests, carries work forward with visible progress, and keeps you in the loop at checkpoints. This isn’t single-turn chat anymore. This is sustained, goal-oriented execution.
Agent 365, generally available in May 2026, is Microsoft’s inventory, monitoring and governance layer for agents that are brought under its management. The Microsoft 365 E7 Frontier Suite bundles it with Microsoft 365 Copilot, Entra, Defender, Intune and Purview capabilities. The control plane only works for agents you discover and onboard; it is not automatic visibility into every script that happens to call a model.
- Chat — "help me write this"A user steers a conversation and turns the answer into action. Context can persist within the chat, but the user remains the orchestrator.
- Code — "build it for me"The delegated job becomes the unit of work. The agent reads the repository, plans, uses tools, runs the tests and opens the pull request.
- Claw — "just handle it"A persistent agent running on hardware you control, connected to messaging channels and acting from configured messages, schedules or events.
The pattern: who carries the work?
The three labels describe responsibility more usefully than chronology:
Chat: the person steers a conversation and applies the result. Think Microsoft 365 Copilot Wave 1.
Code: the person delegates a bounded job; the agent plans and acts. Think Foundry Agent Service, Copilot Studio or GitHub Copilot coding agent.
Claw: a persistent agent remains connected to channels, triggers and tools. Think OpenClaw or a Cowork-style agent, governed through a control plane such as Agent 365.
As more work moves from the user into the system, the governance target shifts from prompts and outputs to identities, tools, data paths and actions.
What this means for you
The governance boundary has moved
Chat systems already carried risks: sensitive data disclosure, hallucinated advice, prompt injection and intellectual-property leakage. Tool-using agents add a larger action surface. They authenticate to systems, cross data boundaries and may communicate with other agents. The possible consequence is no longer confined to a bad answer.
This is why Microsoft is investing in agent governance. Agent 365, Entra, Purview and Defender provide parts of the identity, data and threat-control story, with coverage depending on the agent surface and integration. I mapped those dependencies in Outside the Model: Governing Agents on the Microsoft Stack.
Your platform strategy needs an agent layer
Most organizations have a cloud strategy and a data strategy. You now need an agent strategy. That means deciding where agents get built (Foundry for pro-dev, Copilot Studio for business users, Power Platform for workflow automation), how they interoperate (A2A protocol, MCP), who can deploy them, and how they’re monitored. Microsoft’s stack gives you the pieces. But assembling them into a coherent operating model is your job.
Open source is setting patterns
OpenClaw moved from a personal open-source project into enterprise conversations remarkably quickly, and NVIDIA began building around the pattern. Enterprise platforms are absorbing the same ideas: persistent state, local tools, messaging channels and event-driven work. Security teams should study those mechanics before employees introduce them informally.
Model choice is becoming less permanent
Wave 3 also makes Microsoft’s model diversity explicit. Copilot Cowork uses Anthropic technology, while Foundry exposes models from several providers. That does not make model quality irrelevant, but it makes a permanent single-model bet harder to defend. Keep the model boundary replaceable and govern capabilities at the tool and identity layers.
The agent-operated enterprise
The next stage is likely to involve specialised agents that hand work to each other and share context. That possibility is exactly why the three-pattern model matters. A fleet cannot inherit more authority than the organisation knows how to govern.
Microsoft describes that direction as a “human-led, agent-operated” organisation. Copilot is the interface, Foundry and Copilot Studio are building surfaces, and Agent 365 is the control plane. It is a useful architecture, provided the controls are tested against the agents and tools you actually run.
Start by inventorying which of the three patterns already exists in your organisation. Strengthen the data-governance foundation, pilot bounded agents and study the open-source tools your employees can run without a procurement project. Then ask the harder question: what changes when software can operate alongside your people rather than merely advise them? The July 2026 model-evaluation incident shows why that boundary matters.
The claw is here. The question is whether you’ll ride the lobster, or watch it from the shore.
Sources: OpenAI’s original ChatGPT announcement describes dialogue and follow-up questions from launch. Microsoft detail from the March 2026 Frontier Suite announcement, Foundry Agent Service general availability and Microsoft Agent 365 documentation. Open-source product detail from the OpenClaw repository and NVIDIA’s NemoClaw announcement. Product availability reflects documentation checked on 25 July 2026.