Agentic loop
About AI in the development world
different types of AI use in development
First things first: there is a lot of misinformation and misunderstanding in these topics. Because it is a new and evolving technology, it happens a lot that people are misusing or misinterpreting new words and phrases.
prompting / AI assisted development
This is the most basic form of using the AI for coding tasks. Asking simple questions about the source, generating comments, generating commit messages, code completion. Very basic, very simple, helps a little, but not much.
vibe coding
The (currently) most dangerous way to use AI.
Have a prompt, tell it what you want, let it do it’s thing. Uncontrolled, unprepared, just raw “do whatever”.
The thing where people without technical understanding produce code that is unmaintainable, unsecure and unreliable. If for whatever reason that code then goes anywhere near productivity this is a risky gamble and could lead to problems.
agentic coding
The senior developer’s vibe coding. You have agents (prompts with flavour) with tools (usually MCPs and control over your IDE) that pursue a certain goal.
This can be enhanced by allowing an agent to start other agents (with other flavours and tools) for specific tasks.
agentic loop coding
This is what i do.
You have an “orchestrator” with a very well defined flavour (the agent definition) that analyzes the topic that you give him, collects information about the task, handles specialized sub-agents. Those sub-agents have control-agents that review their work and throw it back if it does not fit.
To improve even more on that, every agent-flavour uses the model best suited for it’s task. Different models for planning, for reviewing, for coding, for security-checks.
The tricky part is to get the agent-system-prompts (in zoo code it’s called “mode”) right and to provide it with skills that improve the quality while at the same time saving tokens.
This is an early version of the agentic loop i designed.
Be aware that one of the key features of this loop is that any review task MUST NOT use the same model-family as the implementation task. That targets planning and dispatch/implementation.
Currently in my setup there are multiple different loops that work together, where the orchestrator decides which loop is the right one for the job.
I will go into detail in a future post.