Beyond autocomplete: building agents you can read, limit and count.
Developers already use AI to write code, and the next skill is building systems where a model picks the next step. Here is what that changes, where it breaks, and how to learn it.
Standup, then a pull request review where half the diff was generated and the tests pass but one branch is never reached. Late morning goes to a feature ticket that says the assistant should look up an account and offer a refund, which turns into an hour of questions about what it may touch. After lunch, a flaky agent run: the trace shows the model picked the wrong tool three steps before anything visibly failed. The day ends with a rerun of the test set and a count.
Coding assistants are now part of the toolchain for many developers. The more interesting shift is that teams are being asked to ship features where a model chooses what happens next: a support agent that decides which account to look up, a pipeline that picks a tool, a bot that drafts and sends. That is a different engineering problem from autocomplete, and most of the hard parts are not about prompts.
Writing code. Assistants draft functions, tests and migrations, and explain unfamiliar code. The review burden moves to you. Generated code is plausible by construction, so a test that passes is the start of the check, not the end.
Building with models. The request from product is often no longer "add a chatbot." It is "have it look up the order, decide if a refund applies, and issue it." That means tool definitions a model can call correctly, output shapes your code can act on, permissions that stop the model doing what it should not, and a human approval step before anything it cannot undo.
Debugging. A failing agent run does not throw a stack trace at the bug. It produces a trace of steps, and the first wrong step is often several turns before the visible failure. Reading a trace forward and fixing the first wrong step is a skill.
Evaluation. The same prompt with the same input gives different answers on different runs. "It worked when I tried it" is one sample. Judging an agent means running its test set several times and counting how many runs came out right.
AI Agents Fundamentals is for developers who write Python and have called a model from their own code. You tell an agent from a chatbot or fixed workflow by what each does at a surprise, and work out whether a task is worth an agent from its cost and failure rate. You write a system prompt in five parts and test it against fixed edge cases, design the agent's tools and permissions with a person approving any step it cannot undo, and turn a Python function into a tool with a description and argument schema. You build the agent without a framework hiding the steps, judge it by running its test set several times and counting correct runs, and read a trace forward to find the first wrong step. About 9 hours, 20 lessons, from $9 a month.
If you lead a team that uses assistants but has never measured the check, the free Start Here course is a two-hour exercise in timing one task by hand and with an assistant and writing a verdict on it.