Most AI agents today are like instant noodles: quick to put together, good enough for a demo, but nowhere close to Michelin-star quality. You paste a prompt into your favorite LLM, tweak a few words, and hope it behaves. The result works… until it doesn’t. Agentic Engineering Weekly is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. In this episode, we’ll break down how to move from that “doom word smithing” phase into a disciplined, prompt optimization workflow for agentic applications. Using a ramen kitchen analogy, we’ll walk through levels of maturity—from manual tinkering to automated optimization loops with specialized “chef” agents. Segment 1 – The impossible triangle and bland AI Every agent builder runs into the same impossible triangle: speed, cost, and accuracy. You can usually optimize for two, but rarely all three at once. * Fast and accurate often means expensive models or long prompts. * Fast and cheap usually means you sacrifice reliability. * Cheap and accurate can become painfully slow. Most teams respond with manual prompt engineering—endless edits, extra instructions, more guardrails—which is basically throwing random ingredients into your ramen pot and hoping it tastes better. Without a recipe or a taste test, it’s guesswork. Segment 2 – Level 1: Adding evals (the taste test) The first real upgrade is evals: defining what “good” looks like and measuring it consistently. This is your taste test. You move from “I think this looks better” to concrete checks: did the agent follow constraints, reason correctly, and produce the right structure on a labeled set of examples. Evals turn subjective tinkering into an objective feedback loop. Once you have that tasting system in place, a key question appears: instead of manually changing the recipe, can another agent do that for you? Segment 3 – Level 2: The brute-force robot chef At Level 2, you bring in a meta-optimizer: an AI that generates many variants of your base prompt, runs them through your eval, and ranks them. Think of this as a very fast, not-very-creative robot chef. * It produces hundreds or thousands of small variations of your prompt. * Each variation is evaluated on your test set. * The best-performing prompts are promoted; the bad ones are discarded. This brute-force search already outperforms manual editing because you explore a much larger design space, guided by metrics instead of vibes. But it’s still inefficient—you’re throwing a lot of spaghetti, or ramen, at the wall. Segment 4 – Level 3: The Michelin kitchen of optimizers At Level 3, the “kitchen” becomes a team of specialized optimizer agents, each with a different role. 4.1 Evolutionary “bake-off” chef Evolutionary algorithms treat prompts like a population of recipes competing in a tournament. * You start with many candidate prompts. * The best ones become “parents”, combining and mutating into new “children” prompts. * Over generations, performance improves as weak prompts are eliminated. Advanced methods like GEPA break the prompt into components—system message, task description, examples—and optimize each part separately before recombining them. 4.2 The Doctor-Chef: HRPO HRPO, or Hierarchical Reflective Prompt Optimization, behaves like a diagnostic chef. Instead of spewing random variants, it analyzes where the agent fails: did it hallucinate, misread constraints, or mis-handle edge cases. It then creates targeted patches that address each failure mode, resulting in surgical edits rather than wholesale rewrites. 4.3 The Sommelier: parameters and few-shot selection Sometimes the recipe is fine; the oven temperature and plating are off. * A parameter optimizer tunes settings like temperature and top_p to get the best trade-off between creativity and reliability for the same prompt. * A few-shot Bayesian optimizer finds the minimal, most informative examples to include, so the agent learns the task with fewer tokens and better generalization. Together, these optimizers give you a rich toolbox instead of a single “better prompt” button. Segment 5 – Choosing the right “chef” and wrapping up Which optimizer should you use? It depends on your constraints and failure modes. If you just need a quick lift on a simple task, a basic meta-prompt optimizer might be enough. On more complex benchmarks, hierarchical approaches like HRPO often emerge as strong all-round performers, especially when you care about fine-grained failure analysis and robust improvements over a baseline prompt. The big shift is mental: you stop thinking in terms of one perfect prompt and start thinking in terms of a continuous optimization pipeline. If your agents still feel like instant noodles, it’s time to invest in evals, optimizers, and a proper kitchen. With the right systems, you can reliably move from quick prototypes to Michelin-star-grade agents that hold up in production. Agentic Engineering Weekly is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. Thanks for reading Agentic Engineering Weekly! This post is public so feel free to share it. Get full access to Agentic Engineering Weekly at shreyassk.substack.com/subscribe
