automating intentions we don't yet understand
October 2026
The simplest picture of an AI organisation looks remarkably like the existing one. There is a research agent, a planning agent, an many more agents for everything. They receive assignments, send updates and pass things along. Someone has drawn the same organisation chart with different occupants. It is the logical place to start. The same task goes in, a cheaper or faster result comes out using AI, and the rest of the organisation continues as before. Existing roles provide boundaries, and familiar workflows make an unfamiliar system easier to supervise. But they also carry assumptions about the people those workflows were built around. Replacing the person with an agent doesn’t automatically make the surrounding structure appropriate.
This thinking reflects human limitations. An atomic swap of human reasoning for model inference assumes that all of this can remain roughly where it is. AI adds the possibility of executing human interpretations of value faster, cheaper and more consistently. It doesn’t increase the probability that these interperations where the best way to complete such task in first place. This makes workflow redesign more difficult than identifying every step an agent can perform. A weekly status meeting might be an expensive way to reconstruct information already present in several systems. If that state becomes available directly, the first response is to replace the meeting with an agent-written update. But if relevant changes can be detected and acted on continuously, the weekly cycle itself becomes questionable. Feedback can alter work while it is happening, rather than wait for the next scheduled review. The design problem becomes deciding which changes warrant a response and whose attention they need.
Moden agentc thiking should not require recreating an office inside the agent, but it does require coordination designed for the system that actually exists. There is still a lot to learn about what those arrangements look like when they operate continuously, across organisations and over long periods. Copying human workflows is a reasonable transitional method, but a poor default for long-term design. It provides places to inspect the work and familiar points at which to intervene. The mistake would be treating that first translation as the destination.