An Agent is an Office
Here's a visual analogy for you to help you unpack the experience of turning a large language model into an agent.
by Joshua Crowley
There's been a lot of discussion about agents this year, and much of it has felt hollow. Engineers describe implementation details—frameworks, orchestration patterns, tool-calling mechanisms—but rarely what an agentic experience actually feels like. Yet agents are real. I've spent countless hours working with them, watching them think, make decisions, and accomplish tasks I couldn't easily do alone.
The most obvious analogy for large language models is human—after all, we interact with them through natural language. But humans and AI are fundamentally different, and conflating the two causes confusion. The formless nature of these models means we need to be intentional about how we adapt them into our workflows, avoiding unwanted side effects that weren't meant to be part of the experience we're creating.
So let's explore a different analogy—one that I think fits remarkably well. If a language model is a room you inhabit for a conversation, then an agent is an office with capabilities—a workspace that can act, remember, and grow more powerful as you give it the right tools. It's perhaps not the most thrilling analogy for what's been hyped all year, but I think the idea of an office really gets you thinking about the kinds of agents you want to work with. So follow me as I expand on this idea.
The Basic Office
Large language models are built to recreate the experience of conversation—a user and an assistant exchanging messages, each response predicted from what came before. That's the foundation, but it's not enough for an agent. What transforms this conversational space into an agent is what you furnish it with: tools, time, and memory. These capabilities turn passive dialogue into active problem-solving.
The Clocks
Clocks line the walls, symbolizing the agent's ability to take time—to reason, iterate, and think deeply. An agent doesn't rush to respond. Instead, it generates tokens to furnish the room with thought, considering approaches, evaluating options, refining solutions. Time transforms quick reactions into careful deliberation, shallow attempts into sophisticated strategies.
The Bureau
The bureau is where the agent writes and edits files. This isn't just reading; it's creation and modification. The agent drafts documents, refines code, restructures data. The bureau represents direct manipulation of the world outside the room, transforming thought into tangible artifacts that persist beyond the conversation.
The Filing Cabinet
Filing cabinets line the walls—a shared mental model both you and the agent understand. Artifacts, skills, and knowledge stored and retrieved on demand. The file system becomes a folio of capabilities, a structured memory that transcends the limitations of context windows. The agent doesn't just remember; it organizes, categorizes, and systematically accesses what it needs.
The Computer
Giving an agent access to a computer proves remarkably powerful. The agent can browse the internet, pulling fresh information into the room. It can write scripts and execute them, bringing computational results back as furnishings for further reasoning. The computer is a window to capabilities beyond language—calculation, automation, real-time information retrieval.
Office buildings
Once you have a functional agentic room, the question becomes: can we stack and combine them? This is the active frontier of AI experimentation—connecting rooms together in increasingly sophisticated architectures. Some believe composition is where the real power emerges. Others are still exploring what works.
Workflows
Rooms connect through stairways—pathways representing specific workflows. An agent in one room completes its work, then passes the results up or down the stairs to another room with different capabilities, different context, different focus. The agent decides when to traverse these stairs, moving work through a tower of specialized spaces, each contributing its unique strengths to the overall task. OpenAI's AgentKit is exploring these patterns, creating frameworks for agents to work together in structured ways.
Sub-agents
Scale this further and you have an apartment building—multiple sub-agents living separate but connected lives. Each room has its own personality, its own focus, its own responsibilities. They're distinct yet cooperative, sharing the work of accomplishing a task too complex for any single room. Different tenants specializing in different skills, collaborating through shared hallways and common spaces. Claude Code's subagents embrace this approach, letting you create specialized agents that work together seamlessly.
AGI
There's intense focus on scaling up to entire cities—vast networks of specialized agents across data centers, each room representing a different facet of enormously complex tasks. Labor divided across neighborhoods, districts, entire metropolitan areas of computation. This is the frontier, the path to tackling problems that seem impossibly large today. Cities of agents working on humanity's hardest challenges. Could this be the path to AGI? As these cities grow more sophisticated, more interconnected, more capable of self-organization and emergent behavior, we might find that general intelligence isn't a single breakthrough but the natural consequence of enough specialized rooms learning to work together. Google's AI co-scientist research hints at this future, where agent systems accelerate scientific discovery at unprecedented scales.
Your Personal Office
Recognizing the shift from software applications to agentic workspaces is a big opportunity for individuals. At the moment, carefully curating what data you bring in, planning what models you use, and what capabilities you give those models is a new skill—and it's going to radically change how you work in a very personal way.
The Ransacked Office
As people experiment with consolidating their personal lives into agentic workspaces, these offices become prime targets for hackers and scammers. An office that holds your emails, calendar, financial records, and personal documents is a treasure trove worth breaking into. The more we centralize our digital lives into these powerful assistants, the more attractive they become to those with malicious intent. Security isn't just about what your agent can access—it's about who else might gain access to your agent.
From My Office
This article itself came from my office. I worked with Cursor—an agent that helped me think through these ideas, building on a previous article I'd written. We went back and forth, discussing the metaphor, refining the concepts. I asked it to generate the image generation scripts, then to help me lay out the images you've been scrolling through. The agent pattern has definitely arrived this year, and it starts with understanding the offices we're building, and feeling some ownership over how we work within them. I'm enjoying getting comfortable in my office, and I hope you are too.
