JANUARY 2026
What I Learned Building Agentic AI Systems
Lessons from taking agentic platforms to production in regulated manufacturing — why the graph matters more than the model, and where autonomy actually pays.
The pitch deck version of agentic AI is an autonomous agent wandering a toolshed, doing your job while you watch. The production version I have shipped looks nothing like that. Here is what building pharmaceutical recipe-authoring and validation platforms actually taught me.
1. The graph is the product; the model is a component
The moment our workflow had a typed, persisted state machine — draft, validate, retrieve, review, refine — quality stopped being a prompt-engineering story. Every serious improvement came from restructuring the graph: adding a validation node, splitting retrieval into tiers, moving a decision from the model into code.
If you are “prompt engineering” inside a 400-line system message, you have not designed a workflow yet.
2. Human-in-the-loop is an architecture, not a checkbox
HITL gets discussed as an ethics sticker. In practice it is the mechanism that makes adoption happen: domain experts accept what they can inspect, pause, and override. Designing the review experience — what the expert sees, what they can change, how their decision is captured — mattered as much as the model’s accuracy. The gate is also your audit trail; treat it as a feature of the state machine, not an interruption of it.
3. Autonomy is a dial, and production wants it low
Every decision we moved from “model judgment” to “deterministic code” reduced failure investigations. Routing by rule where a rule exists; reserving the model for genuine judgment calls. The irony of agentic AI: the more reliable the agent, the less it decides.
4. Tools are contracts; treat them like APIs
Schema-validated tools with small, purposeful surfaces beat kitchen-sink tools every time. We standardized our tool layer on MCP precisely because agents kept changing and the enterprise systems underneath did not. The description of a tool is a prompt — vague descriptions produce vague calls.
5. Evaluation has to be per-node
End-to-end averages hid a retrieval regression for weeks. Node-level metrics made it visible in minutes. In agentic systems, evaluate every node; the trace is the spec.
6. The bottleneck moves — design for where it goes
We removed months of assembly work, and expert review time became the constraint. That is a success, but it was a surprise: systems that remove a bottleneck move the pressure somewhere else. Watch where, and design for the next constraint.
What I would tell teams starting now
Write the workflow without AI first. If you cannot draw the state machine, an agent will not save you. Then keep the autonomy low, the gates human, the state persisted, and the evaluation honest. That is the whole discipline.