<?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: Arthur Jacquemin</title>
    <description>The latest articles on DEV Community by Arthur Jacquemin (@arthurjacquemin).</description>
    <link>https://dev.to/arthurjacquemin</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%2F4242705%2F27f7b1f8-4c3e-4da8-996f-6c8fc8284419.jpg</url>
      <title>DEV Community: Arthur Jacquemin</title>
      <link>https://dev.to/arthurjacquemin</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/arthurjacquemin"/>
    <language>en</language>
    <item>
      <title>Event Sourcing for AI agents: kill the process, resume, replay</title>
      <dc:creator>Arthur Jacquemin</dc:creator>
      <pubDate>Fri, 13 Nov 2026 15:06:55 +0000</pubDate>
      <link>https://dev.to/arthurjacquemin/event-sourcing-for-ai-agents-kill-the-process-resume-replay-4apg</link>
      <guid>https://dev.to/arthurjacquemin/event-sourcing-for-ai-agents-kill-the-process-resume-replay-4apg</guid>
      <description>&lt;p&gt;In &lt;a href="https://comparetiers.com" rel="noopener noreferrer"&gt;comparetiers&lt;/a&gt;, I built a SaaS price-comparison agent that makes eight model calls and runs three tools.&lt;/p&gt;

&lt;p&gt;But what if the process dies halfway? Everything is lost!&lt;/p&gt;

&lt;p&gt;So I changed its state to be an event journal, so:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;you can stop this agent at any moment and run &lt;code&gt;resume&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;you can replay a whole run without calling the model&lt;/li&gt;
&lt;li&gt;the journal is an audit trail, line by line&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The code is written in TypeScript and I use qwen3:8b with Ollama.&lt;/p&gt;

&lt;h2&gt;
  
  
  Initial problem: state in local variables
&lt;/h2&gt;

&lt;p&gt;My first version of the agent loop looked like this:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="kd"&gt;let&lt;/span&gt; &lt;span class="n"&gt;turn&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="no"&gt;MAX_TURNS&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;8&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;messages&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;system&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;user&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="k"&gt;while&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;turn&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;=&lt;/span&gt; &lt;span class="no"&gt;MAX_TURNS&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;decision&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nf"&gt;callModel&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;messages&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;result&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nf"&gt;runTool&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;decision&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;toolCall&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

  &lt;span class="n"&gt;messages&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;push&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;decision&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;result&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="n"&gt;turn&lt;/span&gt;&lt;span class="o"&gt;++&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;All the state (messages, turn counter, amounts seen so far) lives on the stack. There's no resume point and no reliable history.&lt;/p&gt;

&lt;h2&gt;
  
  
  The model: one event per step
&lt;/h2&gt;

&lt;p&gt;I represented a run as a &lt;a href="https://jsonl.co/guide/what-is-jsonl" rel="noopener noreferrer"&gt;JSONL&lt;/a&gt; file. Each line is an event, validated by &lt;a href="https://zod.dev" rel="noopener noreferrer"&gt;zod&lt;/a&gt;.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;run_started -&amp;gt; turn_started -&amp;gt; model_decision -&amp;gt; tool_called -&amp;gt; tool_result -&amp;gt; ... -&amp;gt; run_ended
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Plus &lt;code&gt;model_failed&lt;/code&gt;, &lt;code&gt;tool_error&lt;/code&gt;, and a &lt;code&gt;run_resumed&lt;/code&gt; on each resume. A real excerpt:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"seq"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="nl"&gt;"kind"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="s2"&gt;"run_started"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="nl"&gt;"question"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="s2"&gt;"How much is Notion Plus for 5 seats, billed monthly?"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="nl"&gt;"maxTurns"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="mi"&gt;8&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="nl"&gt;"model"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="s2"&gt;"qwen3:8b"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"seq"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="nl"&gt;"kind"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="s2"&gt;"turn_started"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="nl"&gt;"turn"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"seq"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="mi"&gt;3&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="nl"&gt;"kind"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="s2"&gt;"model_decision"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="nl"&gt;"turn"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="nl"&gt;"toolCall"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"name"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="s2"&gt;"get_plans"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="nl"&gt;"arguments"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"product"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="s2"&gt;"notion"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="nl"&gt;"usage"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"promptTokens"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="mi"&gt;514&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="nl"&gt;"outputTokens"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="mi"&gt;22&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="nl"&gt;"durationMs"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="mi"&gt;9272&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The schema is a discriminated union on &lt;code&gt;kind&lt;/code&gt;. Every line read back is parsed again, so a corrupt journal fails loudly instead of producing a strange state.&lt;/p&gt;

&lt;p&gt;Important: &lt;code&gt;turn_started&lt;/code&gt; is written BEFORE the model call so a turn cut off mid-call shows up in the journal and counts against the budget. Crashing doesn't buy extra turns.&lt;/p&gt;

&lt;h2&gt;
  
  
  The durability rule: write, fsync, then act
&lt;/h2&gt;

&lt;p&gt;So an event is written and synced to a storage before the agent acts on it:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;closeSync&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;fsyncSync&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;openSync&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;writeSync&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="s2"&gt;"node:fs"&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="k"&gt;export&lt;/span&gt; &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;appendEvent&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt;
  &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;path&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;string&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt;
  &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;event&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nc"&gt;Event&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="k"&gt;void&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;fd&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;openSync&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;path&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"a"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
    &lt;span class="k"&gt;try&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
      &lt;span class="nf"&gt;writeSync&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;fd&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sb"&gt;`${JSON.stringify(event)}\n`&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
      &lt;span class="nf"&gt;fsyncSync&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;fd&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="k"&gt;finally&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
      &lt;span class="nf"&gt;closeSync&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;fd&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;runId&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"r-1"&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;append&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;appendEvent&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"./runs/${runId}.jsonl"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="nf"&gt;append&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;seq&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;kind&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"turn_started"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;turn&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="nf"&gt;append&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;seq&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;4&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;kind&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"tool_called"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;turn&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;toolCall&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;name&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"get_plans"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;arguments&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;product&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"notion"&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;tool_called&lt;/code&gt; is on disk before the tool runs, &lt;code&gt;tool_result&lt;/code&gt; is on disk before the next model call, so on resume there are only a handful of cases, depending on the last event:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;last event&lt;/th&gt;
&lt;th&gt;
&lt;code&gt;resume&lt;/code&gt; does&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;tool_called&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;runs the tool again (all tools are read-only)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;code&gt;model_decision&lt;/code&gt; with a tool&lt;/td&gt;
&lt;td&gt;writes &lt;code&gt;tool_called&lt;/code&gt;, runs the tool&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;code&gt;turn_started&lt;/code&gt;, &lt;code&gt;model_failed&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;new turn; the lost one is already counted&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;a result, or &lt;code&gt;run_started&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;applies the stop rules, then a new turn or &lt;code&gt;run_ended&lt;/code&gt;
&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;run_ended&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;refuses (exit 3)&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;There's no other case to handle.&lt;/p&gt;

&lt;p&gt;A precision for the first row: running a tool again is only safe because all tools are read-only. The guarantee is "at least once", not "exactly once". A tool with a side effect (a payment, sending an email) would need an idempotency key so it doesn't run twice. &lt;/p&gt;

&lt;h2&gt;
  
  
  State is &lt;code&gt;fold(events)&lt;/code&gt;
&lt;/h2&gt;

&lt;p&gt;The agent's current state is a pure fonction of the journal:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="k"&gt;export&lt;/span&gt; &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;fold&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;runId&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;string&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;events&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="k"&gt;readonly&lt;/span&gt; &lt;span class="nc"&gt;Event&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nc"&gt;AgentState&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;It's deterministic: same events = same state. Start, resume, the &lt;code&gt;list&lt;/code&gt;command, replay,... all go through it. There's no second copy of the state that could drift. That's the core idea of event sourcing, and it carries over to an agent unchanged.&lt;/p&gt;

&lt;p&gt;Two edge cases need handling:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Torn line&lt;/strong&gt;: a crash during a write leaves an incomplete last line. It's dropped with a warning, and the file is rewritten without it (temp file, then &lt;code&gt;rename&lt;/code&gt;) before the next append. An invalid line anywhere else is a hard error that names the &lt;code&gt;seq&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Lock&lt;/strong&gt;: &lt;code&gt;&amp;lt;runId&amp;gt;.lock&lt;/code&gt; holds the pid driving the run. If that pid is alive, &lt;code&gt;resume&lt;/code&gt; refuses (exit 4). If it's dead, which is what a &lt;code&gt;SIGKILL&lt;/code&gt; leaves behind, the lock is taken over and the old pid is recorded.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Replay: the run without the model
&lt;/h2&gt;

&lt;p&gt;Replay runs the same loop, but with a &lt;code&gt;decide&lt;/code&gt; function that returns the decisions recorded in the journal instead of calling the LLM. By default, tool results also come from the journal. With a flag &lt;code&gt;--live-tools&lt;/code&gt;, the tools actually run against the current catalog, and a divergence means the data has changed since the run. Every event produced is compared to the journal (kind and content, ignoring timestamps and token usage), and replay stops at the first divergence.&lt;/p&gt;

&lt;p&gt;That catches anything the code drifts on: &lt;code&gt;tool_called&lt;/code&gt; arguments that no longer match the decision, a changed stop rule, a final result that no longer leads to the recorded &lt;code&gt;run_ended&lt;/code&gt;. It's a regression test built from real runs.&lt;/p&gt;

&lt;p&gt;What it doesn't catch with recorded results: a &lt;code&gt;tool_result&lt;/code&gt; edited by hand, since it's replayed as-is.&lt;/p&gt;

&lt;h2&gt;
  
  
  What it costs, what it buys
&lt;/h2&gt;

&lt;p&gt;The journal only grows: on long runs, you'd want snapshots so a resume doesn't fold everything from scratch. Changing an event's schema breaks reading older journals, so events will need versioning. And resume only holds for tools that are safe to run twice.&lt;/p&gt;

&lt;p&gt;In return, a crash costs nothing anymore: &lt;code&gt;resume&lt;/code&gt; picks up from the last event without paying again for the turns already done. And every recorded run becomes a free regression test, replayed without a single model call.&lt;/p&gt;

&lt;p&gt;Repo: &lt;a href="https://github.com/Kazdan1994/event-sourcing-agent" rel="noopener noreferrer"&gt;https://github.com/Kazdan1994/event-sourcing-agent&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>typescript</category>
      <category>llm</category>
      <category>eventsourcing</category>
    </item>
  </channel>
</rss>
