84% of developers use or plan to use AI tools. 46% distrust the output.
Most developers use or plan to use AI, and more distrust its output than trust it. Learn to build an agent you can read, limit and count, and to automate the ops glue around your team.
In Stack Overflow's 2025 survey of more than 49,000 developers, 84% use or plan to use AI tools in their work and 51% of professional developers use them daily. Trust runs the other way. 46% distrust the accuracy of the output, 33% trust it, and the top frustration, named by 66%, is answers that are almost right but not quite. 45% said debugging AI-generated code takes more time. The Census Bureau's business survey puts AI use in the information sector at 39.7% as of May 2026, about double the national 19.8%.
So the field is past "should we use it" and into "how do we know when it is wrong."
Output you can act on. A model returns a string. Before anything downstream uses it, the shape is enforced by the API, the values are validated, and the failure path is defined. The agents course spends a chapter on this: the same question giving two answers, the citation that was not a real case, making the API hold the shape, and whether a count going up meant anything improved.
Permissions and blast radius. OWASP's 2025 Top 10 for LLM applications lists prompt injection first, with sensitive information disclosure, improper output handling and excessive agency also on the list. The agents course has you design the tools, the permissions and a human approval step for anything the agent cannot undo, and asks what a read-only agent can still lose and whose rows the model just read.
Evaluation by count. Run the test set several times and count the correct runs. Then read the trace forward to the first wrong step and fix it there.
Cost. Every model step has a price and a failure rate. The course has you decide whether a task deserves an agent by working both out before you write it.
Software developers (BLS projects 10% growth from 2025 to 2035, median pay of $135,980 in May 2025, and about 106,100 openings a year including QA roles), QA analysts and testers, DevOps and platform engineers, developer advocates, engineering managers and technical product managers.
If you write Python and have called a model from code, AI Agents Fundamentals is about 9.5 hours in which you tell an agent from a chatbot or a fixed workflow by what each does at a surprise, write a system prompt in five parts and test it against fixed edge cases, turn a Python function into a tool with a description and argument schema, write the loop and the trace yourself with no framework hiding the steps, and judge the agent by running its test set several times and counting correct runs.
If the work is ops glue rather than code, the automation course builds one intake on Zapier, Make and n8n, checks the record each one wrote, routes each item to one owner with a fallback, and adds one priced, capped model step. A team's Zaps built on one person's account stop when that person leaves, and the course has a lesson on exactly that.
Colleagues outside engineering can start with the free Start Here course, which has them time one real task with an assistant and size the check to the stakes.