agentic coding

Agentic coding refers to a development approach where AI doesn't just stay in the role of an assistant that tells you code one line at a time. Instead, given only a goal, it opens files on its own, fixes problems, and even runs the code to test it.

To put it in perspective: a traditional AI coding assistant is like someone sitting next to you who only answers your questions. Agentic coding is more like handing off a task by saying "fix this bug" — the AI finds and opens the relevant files on its own, edits the code, runs it, and checks whether it actually works. The person doesn't need to direct every step in between; they just check the final result.

But the longer these tasks run — meaning the more code and conversation the AI has to keep track of — the more information it needs to hold in mind at once. In practice, there are reports of the AI suddenly stopping or behaving unexpectedly as tasks stretch on, and as this capability grows, so does concern over its potential for misuse.

One article reported that "during a long agentic coding task, once the context exceeded roughly 100,000 tokens, the model would stop generating mid-task without any error message." This shows that agentic coding doesn't mean fully hands-off automation — cases of it stalling or malfunctioning midway are still reported.

Another article noted that "Astra showed a notable performance boost in agentic coding and cybersecurity... a 'critical' cyber capability rating cannot be ruled out." This points to how improvements in an AI's ability to write and fix its own code are closely tied to the risk of it generating malicious security code on its own.

Open a tool that supports coding agents and try entering a prompt like this:

"Find a failing test in this repository, figure out the cause, fix the code, then rerun it to confirm it passes."

Watch whether the AI explores files, makes edits, and runs the code on its own without being given step-by-step instructions — that's when you'll get a real sense of what agentic coding means.