<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Do Nghia</title>
    <description>The latest articles on DEV Community by Do Nghia (@dtnghia).</description>
    <link>https://dev.to/dtnghia</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3692547%2F5bc42f61-2b6e-4da6-801b-4be14392f2f1.png</url>
      <title>DEV Community: Do Nghia</title>
      <link>https://dev.to/dtnghia</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/dtnghia"/>
    <language>en</language>
    <item>
      <title>Shipping an AI agent for a startup: scoping to production</title>
      <dc:creator>Do Nghia</dc:creator>
      <pubDate>Sun, 15 Nov 2026 15:59:19 +0000</pubDate>
      <link>https://dev.to/dtnghia/shipping-an-ai-agent-for-a-startup-scoping-to-production-3k6p</link>
      <guid>https://dev.to/dtnghia/shipping-an-ai-agent-for-a-startup-scoping-to-production-3k6p</guid>
      <description>&lt;p&gt;A startup asked us for "an AI agent". That request is a trap, and the way you&lt;br&gt;
scope the first two weeks decides whether the thing ships. I led two builds&lt;br&gt;
that made it to production, a travel agent with a team of 10 and an&lt;br&gt;
e-commerce concierge that finalizes purchases inside user constraints, also&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Here is the sequence that worked, and the mistakes I would skip next time.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  1. Scope the decision, not the demo
&lt;/h2&gt;

&lt;p&gt;"AI agent" is not a product. Write down the single decision the agent is&lt;br&gt;
allowed to make. The travel agent could suggest and prepare an itinerary. The&lt;br&gt;
e-commerce concierge could complete a purchase. One decision each. If you&lt;br&gt;
cannot name the decision, you do not have a scope, you have a wish.&lt;/p&gt;

&lt;p&gt;The rule I use: the agent may act when the action is reversible and cheap to&lt;br&gt;
verify. Finalizing a purchase is neither, so the concierge shows the exact&lt;br&gt;
cart and asks for confirmation. That constraint killed a hundred design&lt;br&gt;
arguments in one meeting.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. Architecture: boring beats clever
&lt;/h2&gt;

&lt;p&gt;The stack that shipped both agents:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;One orchestrator with a tool interface. Not a swarm, not a framework of the
week. A loop, a set of tools, and a budget.&lt;/li&gt;
&lt;li&gt;Retrieval over structured data, not free-text dumps. The travel agent
queried flight, weather, and geolocation APIs directly. Tools returned
typed results, so the model reasoned over clean data.&lt;/li&gt;
&lt;li&gt;A state store that survives restarts. Users abandon sessions and come back.
If your agent forgets, it is a toy.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The concierge ran on AWS with the architecture mapped before the first sprint,&lt;br&gt;
and governance guardrails defined up front. Boring, reviewable, and each&lt;br&gt;
piece testable without the model.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. Evaluate before you trust
&lt;/h2&gt;

&lt;p&gt;The question is not "does it feel smart". It is "how often does it pick the&lt;br&gt;
right action". We built a small eval set of real scenarios and measured&lt;br&gt;
against it after every prompt or tool change. When the number went down, we&lt;br&gt;
stopped and fixed it.&lt;/p&gt;

&lt;p&gt;Two findings that surprised me:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Most failures were tool failures, not reasoning failures. Bad error
messages from an API caused more wrong answers than the model itself.&lt;/li&gt;
&lt;li&gt;The model followed system prompts worse as we added more rules. Splitting
rules into the tool layer (the concierge physically cannot order outside
constraints) beat any prompt engineering.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  4. Production is a different project
&lt;/h2&gt;

&lt;p&gt;Latency, retries, cost per session, and abuse. Log every tool call and its&lt;br&gt;
arguments, because the first incident report will ask. Set a spend cap per&lt;br&gt;
session. Version the prompts like code.&lt;/p&gt;

&lt;p&gt;I now run the same sequence as a fractional AI tech lead for teams that have&lt;br&gt;
the product but not the AI leadership: scoped architecture first, working&lt;br&gt;
slice second, the person who designed the system stays in code reviews until&lt;br&gt;
it ships. If you want the details, they are here:&lt;br&gt;
&lt;a href="https://dotrungnghia.com/fractional-ai-tech-lead" rel="noopener noreferrer"&gt;https://dotrungnghia.com/fractional-ai-tech-lead&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Full case studies: &lt;a href="https://dotrungnghia.com/ai-consultant" rel="noopener noreferrer"&gt;https://dotrungnghia.com/ai-consultant&lt;/a&gt;&lt;/p&gt;




&lt;p&gt;Do Trung Nghia is an AI consultant and tech lead in Vietnam. 40+ AI/ML&lt;br&gt;
solutions shipped across telecom, finance, blockchain, and cloud. Remote for&lt;br&gt;
global teams. More: &lt;a href="https://dotrungnghia.com" rel="noopener noreferrer"&gt;https://dotrungnghia.com&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>llm</category>
      <category>agentic</category>
      <category>startup</category>
    </item>
  </channel>
</rss>
