AI & LLMs
What is agentic coding?
Agentic coding is an approach in which an AI model works toward a stated goal on its own: planning the steps, writing code across files, running it, reading the errors, and correcting them without turn-by-turn direction.
New to Bolt? Sign up free.
Technical definition
Agentic coding describes AI systems that pursue a development objective through an autonomous loop rather than a single generation step. The agent decomposes a goal into tasks, selects and calls tools such as file operations, package installation, a terminal, or a browser, executes code, observes the result, and revises its approach based on what it observes.
The distinguishing property is the feedback loop. A non-agentic model produces output and stops. An agent acts, sees the consequence, and acts again.
In plain terms
You state the goal. The AI figures out the steps, does them, checks whether they worked, and fixes what did not.
The difference from ordinary prompting: you are not walking it through the work. It runs the code, sees the error, and handles it.
Why it matters
The loop is what makes AI building practical rather than a demo. A model that writes code it never runs produces plausible code. A model that runs its own code catches the import that does not exist, the type that does not match, and the query that returns nothing.
For someone who cannot read the code, this matters more than anything else in the category. You have no way to spot a subtle error by eye. An agent that tests its own work catches a whole class of failure before it ever reaches you.
The trade-off is oversight. An agent working across many files can make sweeping changes fast, including wrong ones. Small requests, visible checkpoints, and the ability to roll back are what keep that manageable.
