Agents and the work around them
A capable model is a beginning. Useful autonomy also needs context, tools, permissions, and feedback.
Beyond the chat window.
The environment is part of the agent
An agent is easier to understand when we look beyond its model. What can it observe? Which actions are available? How does it find out whether those actions worked? These choices shape the work it can reliably perform.
A demonstration can hide that surrounding structure. A deployed system has to make it explicit. The integrations, exception paths, and evaluation process are part of the product.
The unit of useful work
Instead of asking how autonomous a system is, it may be more productive to ask which unit of work it can finish. A well-defined task with clear feedback is a different problem from a broad objective whose success is hard to observe.
The boundary matters. Handing a task back to a human can be a sensible part of the design when the cost of an error is high or the goal is underspecified.
What to watch
Look for evidence that a system remains useful when the easy cases have run out. How does it recover? Can a person understand its state? Does it improve the full workflow, including review and correction?
The interesting companies may be the ones that make the surrounding work legible enough for both people and machines.
This is an illustrative essay written for the website design preview.
Next essay