Pre-built tools are a dead end for AI agents, and the industry's reliance on them reveals a fundamental misunderstanding of what autonomous systems actually need. The argument laid out in the Plan, Code, Execute framework from Towards Data Science makes this clear: if you give an agent a fixed set of functions, you are capping its potential before it even starts. For anyone building or evaluating agentic architectures, this is not a theoretical concern, it is a practical limit on what your systems can achieve.
Think about what happens when an agent encounters a task its toolkit wasn't designed for. It fails, or worse, it forces a brittle workaround that introduces errors. The alternative is an architecture where agents plan what they need, write the code to accomplish it, and then execute that code on the fly. This is not about giving agents more tools; it is about giving them the ability to make tools. For spreadsheet users, the parallel is direct. You have likely felt the frustration of a formula or pivot table that almost works but requires manual stitching across columns or scripts you have to write yourself. An agent that can create its own spreadsheet functions, dynamically, as needed, turns that friction into a solved problem. The agent becomes a collaborator that adapts to your data, not the other way around.
The practical implication is that the teams and individuals who adopt this self-building approach will leave behind those stuck with static toolkits. Consider your current workflow. How many hours do you spend reshaping data to fit a tool's expectations? How many times have you abandoned an analysis because the available functions couldn't handle the nuance? The Plan, Code, Execute model eliminates that friction by shifting the burden from the user to the agent. It does not require you to become a programmer; it requires the agent to act like one on your behalf. This is where the human-centered promise of AI meets its actual value, not in flashy demos, but in the unglamorous work of making complex tasks feel simple.
Start by asking whether the tools you use today allow the agent to write its own solutions, or whether they handcuff it to a predefined menu. If it is the latter, you are leaving capability on the table. The next evolution of data work belongs to systems that can build the bridge as they cross it.