LLM Agents Are Just Spreadsheets

Do agents eat breakfast before going to work? Does Claude prefer Cornflakes or a parfait? Does ChatGPT enjoy buttered toast? The question may sound absurd. The public is currently roiling in a debate over whether LLMs will bring about a new utopian age where none of us will need to work¹ or turn out to be another crypto. We’re being led to believe it’s only a matter of time before we’re replaced by a genius in a datacenter.²

The models are genies (the magic lamp kind). They grant your wishes (within reason). But they never wish for anything themselves. They don’t wake up and decide what to eat that day. They don’t decide what to work on that day without being provided guidelines, sources of information, and constraints. You could argue I’m not being ambitious enough with what I have agents do. Certainly they could plan what I eat for breakfast. Sure, they can decide what features to add on the product roadmap. But doesn’t that require me giving them a goal? Or giving them guardrails on how to handle my dining preferences? The genie is not out of the bottle.

Don’t get me wrong. Software engineering is serving as the petri dish for the LLM takeover. And indeed, to the casual observer it would appear the public’s fear is well founded. AI can already program better than humans, right? Software engineers used to obsess over tabs vs spaces, measuring test coverage, and pointing stories. Ever since I traded my IDE for Claude Code, those former concerns have faded into the background. The job looks entirely different. I’ve changed my team’s processes to revolve around these new agentic coding capabilities. We continue to invest in finding new and better ways to leverage skills, harnesses, and model updates to improve the quality of what we produce. And yet, our jobs remain stubbornly resistant to the supposed AI takeover. Writing code is no more the entirety of the job than a writer putting pen to paper is all of composing a story.

LLM agents are tools the same way electronic spreadsheets, like Excel, are a tool. Their advent “automated,” or at least dramatically changed, much of the work done at the time. We went from handwritten ledgers to sheets editable on the fly. A 2015 NPR Planet Money episode told the story of the impact of VisiCalc, the first spreadsheet computer program for PCs. It was a predecessor of Lotus 1-2-3 and Microsoft Excel. The episode tells the story of Allen Sneider, a 1970s Boston accountant, discovering the power of VisiCalc. “Sneider says it could take a day - a really boring day - just to make one tweak. … The program could do in seconds what it used to take a person an entire day to do.”³ The electronic spreadsheet brought about sweeping labor automations and changed the shape of jobs. Accountants saw their roles shift as spreadsheets meant they could do more: “What if I hired more employees? What if I charged a little less for my product? What if I borrowed more money? Software made answering questions like these cheap and easy. Accountants became more valuable. They weren’t just adding up numbers, they were thinking creatively about business.” Spreadsheets revolutionized work into the white-collar office jobs we’re so familiar with. Similarly, agents are a new form of work productivity software, ushering in a new normal. This time, we’re all becoming managers.

“Everyone [is] becoming a manager”⁴ resonates with me. It matches my recent experience precisely. Due to a job change, I went from leading an engineering team of seven excellent engineers to being a team of one supported by agents (I’ve since hired two more engineers and would like to hire more!). What has struck me is how similar orchestrating a swarm of agents is to being an engineering manager. The skills of holding work output to the desired quality level are nearly the same. I need to assess the scope of the work we’re planning and make sure we’re not getting ahead of our skis. I have to vet the core assumptions of the architecture we’re shipping and make sure we’re not painting ourselves into a corner when a future feature lands on the roadmap. I have to do this all without reading every line of code. I have to exercise judgment when looking over a technical design doc to make sure the right information was considered, weigh tradeoffs, make decisions, and discern when to dive deeper. All while someone else does the work. Work is becoming more managerial.

There is a weak point in my argument. Perhaps we can replace everyone’s job by creating a cycle of responsibility and deference to authority. In the same way my agents report to me, I report to my boss, my boss reports to the CEO, and the CEO reports to the board or the shareholders. You could run this whole cycle with LLMs, right? But who originates the desire to build the company and the million decisions along the way? Who constructs the scope of what each agent works on, their purview? Who determines how the information is packaged and forwarded down the chain of command? Who decides what feedback is bubbled up from the bottom? Should the org structure be flat? Who decides what factors determine when a new agent should be added to the org chart? All those meta-operations of how a company runs require someone to initiate. It can’t be agents all the way down.

So what is AI? Should we be concerned about our jobs? Should we look forward to sipping margaritas on the beach while the agents toil away in spreadsheets on our behalf? I doubt it. Compared to a year or so ago, my job is unrecognizable thanks to LLMs. It’s undeniable agents have upended software engineering and are poised to do the same to other knowledge worker jobs. But I’m still doing software engineering, probably more of it nowadays. When all is said and done, the agents can’t replace the desires, the wants, a human being brings to work (after eating breakfast, of course) to originate the work for the day.

Further reading