Dev-Diary 1 — Building an AI Agent at Netguru

Dev-Diary #1 — Wiring the Brain

Today’s puzzle: how AI Agent should think.

• Orchestrator — picks the next move.
• Specialist LLMs — experts (fetch data, draft mail, sanity-check output).
• Black boxes — strict in/out ports, fewer side-effects.

AI Orchestrator is our primary Agent. It’s job is to think how to achieve the job and it can use different tools. It is connected to our black-boxes by “tools”. So LLM decides when to use specific blackbox.

What is blackbox? It is the set of functions. Sometimes wired together with simple LLM call, sometimes it is just a function to add numbers. Basically black-box is a fancy tool for our Agent, so it gets only high quality data.

Our first step is to create a functionality to create possible feature list for a system after client request. In classic agentic workflow, agent will think and come up with the answer. But in our case, we already have a lot of processes around feature lists, estimations etc. We need specific format, style and content for our experts for estimations. So agentic workflow will fail. We need structured output -> but this term is related to JSON or YAML formats.

So how to get feature list? We hook up our blackbox as a tool. This blackbox has special prompts to divide project, create modules and then specific features for the system. For now in CSV, as we will import it later to another program.

As you can see, this blackbox has a lot of to do. In chatgpt style conversation at least 3-4 messages to get plan, modules, features etc.

The last step is feedback loop. It is really simple step to ensure quality of outcome. We set up supervisor which is giving feedback for primary agent. It has some guardrails to ensure primary agent is following the task, adding proper content and not making things up.

Check the diagram, it has rough idea how it works. I would like to know how you approach building such systems?