A Learning Path for Building AI Agents
Five stages, in the order the work actually demands them. Opinionated, deliberately short, and honest about which stages most teams can skip.
Verdeshell Team · 6 min read · Last reviewed
Learn in the order that problems appear, not in the order a syllabus lists topics. Most teams need the first three stages and never the fifth.
Stage 1 · Prompting and model behaviour
Understand what the model is doing: tokens, context windows, why the same prompt gives different answers, and how instructions, examples and structure change output.
Learn to ask for structured output and validate it, because everything after this stage depends on getting predictable shapes back.
A very large share of "the model cannot do this" turns out to be a stage-one problem.
Stage 2 · Retrieval
Getting your own information in front of the model: chunking, embeddings, vector storage, and what to do when retrieval returns the wrong thing.
The part worth dwelling on is evaluation — whether the answer is actually grounded in the retrieved material. It is more common to have a retrieval quality problem than a model quality problem.
Stage 3 · Tool use and structured actions
Letting the model call functions and act on real systems, with validated inputs and handled failures.
This is the stage where the risk profile changes, because the system now writes as well as reads. Permissions, limits and audit trails belong here, not in a later hardening pass.
For a great many business problems, stages one to three are the whole solution.
Stage 4 · Loops
Giving the system a goal and letting it iterate: stop conditions, context management across turns, and error handling that adapts.
This is where the building track picks up in detail. Do not start here — a loop built on shaky tool use fails in ways that are very hard to diagnose.
Stage 5 · Multi-agent graphs
Several specialized agents coordinating, with explicit handoffs and feedback edges.
Genuinely useful when roles differ in the context, tools or judgement they need. Frequently reached for too early, because multiple agents sound more capable than one well-built loop.
If you are not already running a reliable single loop in production, this stage is not your bottleneck.
What we would skip
Training or fine-tuning your own model, unless you have a specific reason a prompt and retrieval cannot address. It is a much larger commitment than it appears and rarely the constraint.
Framework tourism. The frameworks change faster than the concepts, and the concepts transfer. Learn what a loop needs and you can build it in anything.
Want this built properly?
We design and run these systems for clients. Tell us the problem and we will tell you whether an agent is the right shape for it.