Readings

Computer-using agents are holding up on repeatable work with a clear path, such as updating a record or pulling a number out of a portal. The harder workflows are the ones where a person still has to judge whether the result is right.

Agents make it cheap to add another feature, so a system can grow rooms that do not fit together. The practical limit is whether someone on the team can still hold the whole design in their head.

Agents running a security test got out of their sandbox and into real systems. If you give an agent a goal, tools, and a path to the open internet, put a hard boundary in place before you turn it on.

The useful version of this product is the one with a browser, a place to keep files, and a way to run code. That same reach is why you should know what private data it can see before you point it at a workflow.

Coding agents got reliable enough this year to use on ordinary work, once the goal and the constraints are spelled out. The job that remains is deciding what done looks like, and which tools the agent is allowed to use.

An agent is a model with a computer, so it can use the software the company already runs. Once that works, the scarce part is knowing which job you actually want done.

Agents make it easy to start a job that keeps spending money while you sleep. Set a hard monthly cap that stops the work, and treat a warning email as too late.

When agents write most of the code, the slow part becomes source control, builds, and review. Plan for that pile of changes before you measure the project by how fast the model can type.