--- title: Three Generations of AI source: https://steadman.ai/newsletters/david/three-generations.html published: 2026-06-03 summary: Chat, Agent, Employee. Three generations of AI that people can actually buy and use today, framed around what you're managing at each level. --- # Three Generations of AI *Chat. Agent. Employee.* Three generations of AI that people can actually buy and use today. Each does more on your behalf than the last, and each asks more of you in return. The verb that runs through all three is *manage*. This is a partnership, not a handover. You are responsible for what the AI produces, every time, whether you're working alone or running an organisation. ## Generation 1: Chat *You manage the execution of a task.* You ask, it answers. One prompt, one response, one decision about whether to use what came back. The conversation is the product. Everything that follows is a step up from this base. ### 1A — Free / Constrained Shorter context, weaker models, no memory. **Limited by design.** Builds foundational skill but sets a low ceiling on what people expect. **Examples:** Free ChatGPT, free Claude, basic Copilot, bundled tools. ### 1B — Pro / Competent Good general-purpose chat applications. **Dramatically more capable** than the free tier: longer context, stronger models, useful features. **Examples:** Paid Claude, paid ChatGPT, paid Gemini. ## Generation 2: Agent *You manage the execution of a complex process.* The AI does work on real things. It holds a task across many steps, uses tools, and pursues an objective with limited intervention. Scheduled tasks and workflows live here too: one good run becomes a process that runs again. **Both flavours below can live inside the same organisation.** The difference is how much deliberate work has gone into safely opening the door. ### 2A — Sandboxed The agent is hemmed in by default. Often the tools are perfectly capable. What's missing is reach into your real systems: no corporate email, no internal databases, no production tools. Sometimes that's how the product ships, sometimes how IT has walled it off, usually both. The isolation is a genuine trade: security and freedom to move outside, in exchange for power inside. **Powerful in its box. The box is the point.** **Examples:** ChatGPT with scheduled tasks, Cowork, Microsoft Copilot's agentic features. The default mode of most enterprise rollouts. ### 2B — Unleashed The agent has real access. Reads files, writes files, calls APIs, finishes pipelines end-to-end. **Achievable inside any organisation, but only with deliberate work.** Identities, permissions, audit trails, sandboxes around the sandboxes. The organisations that succeed here pair IT with the teams doing the work, and set governance that flexes as the tools change rather than freezing on day one. A lone developer reaches this state by default; an enterprise reaches the same capability on purpose, through governance, and never all at once. **Examples:** Claude Code, Cursor, Codex, plus a workflow layer such as Claude Workflows or compound engineering. ## Generation 3: Employee *You manage the execution of a complex role.* The agent has a persistent identity. A name, an email address, an ongoing existence between sessions. You don't open it; you address it, even when you're not at the keyboard. **You hire it, you manage it, you fire it.** You are accountable for its work, the same way you would be for any other employee. A post, not a person: a role your team creates and directs. The point is to give each person an employee of their own, not to replace them with one. ### 3A — Personal Only you address it. It acts as your delegate: it might send emails, call APIs, or work in external systems on your behalf, but it does so as you. The work product lands back with you for delivery. **Every employee starts here.** **Examples:** Single-user Hermes, a personal Manus instance, Anthropic's Claude Managed Agents wired to one account. ### 3B — Team Others address it directly. Colleagues, or clients, email it, ask it questions, expect answers; the conversation bypasses you. **Possible anywhere, with the work done.** The governance widens to cover the audience as well as the agent: who can address it, what it's allowed to commit to, what happens when it's asked something outside its remit. **Examples:** Team-deployed Hermes, OpenClaw with multi-user access, Managed Agents in a shared inbox or Slack. ## Three things to know ### Each generation grows from one into many. One chat becomes many chats open in different tabs. One agent becomes a team of agents working in parallel: sub-agents taking instructions, agents handing tasks to other agents, a small fleet pursuing one goal. One colleague becomes a small staff you delegate across. **This isn't a separate tier. It's something you grow at each level, by choice and under your direction.** Multi-agent orchestration is what Generation 2 and 3 both do as they mature, not a fourth generation in waiting. ### Each tier needs its own kit. A different setup, a different set of permissions, a different set of subscriptions. Those are the breakpoints between generations. Within each tier sits a rich and growing ecosystem of products and configurations; the boundaries *between* tiers are hard. Most organisations allow certain tools and certain versions of certain tools, and stop there. Individuals on home machines are free to use any of it. **Both worlds can reach all three generations.** It just takes more care the further you go. ### It looks like a ladder. It becomes a choice. At first the six levels read as a ladder: you start at 1A and climb. That's the right instinct while you're finding your feet. But fluency changes the question from *how high can I get* to *where should each piece of work sit*. The strongest users spread their work across all six levels at once, picking the lightest tool that does the job. And at every level, the same person stays in the chair: you direct, you check, you own the result. Generation 3 gives each of your people an employee to manage, not a replacement to fear. The further the AI reaches on your behalf, the more it is your judgement, not your absence, that makes the work good. --- An earlier version of this framework (two generations, focused on chat tools vs Claude Code) lives at [Gen 1 vs Gen 2](https://steadman.ai/newsletters/david/gen1-vs-gen2.html). Kept as a historical snapshot. This is an artefact created for David's weekly email. See the others at [steadman.ai/newsletters/david/](https://steadman.ai/newsletters/david/).