<?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: Dmitry G</title>
    <description>The latest articles on DEV Community by Dmitry G (@dmitry_g_f946af88f2ec5a62).</description>
    <link>https://dev.to/dmitry_g_f946af88f2ec5a62</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%2F4254479%2F6f910a98-3afd-498e-b156-ab0393ce714b.png</url>
      <title>DEV Community: Dmitry G</title>
      <link>https://dev.to/dmitry_g_f946af88f2ec5a62</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/dmitry_g_f946af88f2ec5a62"/>
    <language>en</language>
    <item>
      <title>How I get AI agents to finish what they start</title>
      <dc:creator>Dmitry G</dc:creator>
      <pubDate>Thu, 19 Nov 2026 21:03:38 +0000</pubDate>
      <link>https://dev.to/dmitry_g_f946af88f2ec5a62/how-i-get-ai-agents-to-finish-what-they-start-20ij</link>
      <guid>https://dev.to/dmitry_g_f946af88f2ec5a62/how-i-get-ai-agents-to-finish-what-they-start-20ij</guid>
      <description>&lt;p&gt;I asked an AI agent to add a 4.5-meter line to a drawing. It labeled the line “4.5 m,” but the endpoint appeared to land right on the end of the old 3.5-meter line.&lt;/p&gt;

&lt;p&gt;The label was right. The drawing disagreed.&lt;/p&gt;

&lt;p&gt;That’s an easy mistake to spot on a drawing. The version you get in a research report is harder to see. A sentence says a new process cut costs by 30%. You click the citation. The source says costs fell by up to 30% in one pilot. The link works. The sentence doesn’t.&lt;/p&gt;

&lt;p&gt;Both examples point to a question I now ask before handing an agent a task: How will I know the result is right?&lt;/p&gt;

&lt;p&gt;For work that takes several steps, I use three skills. A skill is a set of instructions an agent can reuse. Two of these are mine. The third is by Matt Pocock.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Delivery Roadmap: Decide what counts as done&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;a href="https://github.com/dmitry9g/delivery-roadmap" rel="noopener noreferrer"&gt;Delivery Roadmap&lt;/a&gt; breaks work into small steps and gives each one a check.&lt;/p&gt;

&lt;p&gt;For the drawing, I asked the agent to build a grid, plot points from their coordinates, and verify a known distance before adding the new line. That gave us something to check besides whether the picture looked convincing. The revised drawing was what I’d been looking for.&lt;/p&gt;

&lt;p&gt;For a short research brief, I’d pick the three claims most likely to change a decision and open their sources. If the source says something narrower, narrow the claim. If it doesn’t support the claim at all, take it out. A working link is a good start. So is a ruler with numbers on it.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;a href="https://github.com/dmitry9g/continuation" rel="noopener noreferrer"&gt;Continuation&lt;/a&gt;: Pick up where the work actually stopped&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Continuation helps when a session ends, another agent takes over, or a step gets blocked.&lt;/p&gt;

&lt;p&gt;Say an agent checks two of the three claims before the session ends. A useful handoff says which two it checked, where it found the evidence, and which claim is still open. The next agent can start there. It doesn’t need to research the topic from scratch or assume the third claim is fine because the first two were.&lt;/p&gt;

&lt;p&gt;Keep that note in the roadmap. If you write a new plan every time an agent takes over, you’ll soon have plenty of plans.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Diagnosing Bugs: Make the failure repeatable&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Matt Pocock’s &lt;a href="https://github.com/mattpocock/skills/tree/main/skills/engineering/diagnosing-bugs&lt;br&gt;%0A![%20](https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/fp62cuydt8t9l1ve4ty1.png)" rel="noopener noreferrer"&gt;Diagnosing Bugs&lt;/a&gt; is for when something keeps going wrong and you don’t know why. Find a small test that makes the failure easy to see. Then test possible causes, fix what the evidence points to, and run the check again.&lt;/p&gt;

&lt;p&gt;Suppose a data import drops the last row of a spreadsheet. Try it with three rows. If the third row disappears, you can investigate without importing ten thousand rows every time you have an idea. Once the small test passes, run the original import and check that it actually kept the records. Put both results in the roadmap so the next agent knows what’s been proved.&lt;/p&gt;

&lt;p&gt;Try it on one small job&lt;/p&gt;

&lt;p&gt;Pick a task with a result you can inspect. For a drawing, measure a line. For a report, check a claim. For an import, count the records and look at a few of them. With the three skills available in your session, try this:&lt;br&gt;
Use Delivery Roadmap to plan the work and say what would prove each step is done. Use Continuation to resume from the last verified step. If a bug blocks progress, use Diagnosing Bugs to reproduce it and verify the fix. Keep the evidence in the roadmap. Before calling the job done, check the result the way I’ll use it.&lt;/p&gt;

&lt;p&gt;I want to give an agent a job, go make coffee, and come back without having to reconstruct what happened. I don’t want to hover. It’s a task, not a soufflé.&lt;/p&gt;

&lt;p&gt;The three skills don’t give the agent extra permissions or keep it running in the background. They give me something more useful than a confident “done”: a check I can repeat.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>automation</category>
      <category>agents</category>
      <category>performance</category>
    </item>
  </channel>
</rss>
