<?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: DEIXEI</title>
    <description>The latest articles on DEV Community by DEIXEI (@deixei).</description>
    <link>https://dev.to/deixei</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%2F4245783%2Fd497ce2a-ee11-479e-bac7-b9ba810b1560.png</url>
      <title>DEV Community: DEIXEI</title>
      <link>https://dev.to/deixei</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/deixei"/>
    <language>en</language>
    <item>
      <title>Your coding agent is privileged automation: a workstation baseline for Claude Code</title>
      <dc:creator>DEIXEI</dc:creator>
      <pubDate>Tue, 17 Nov 2026 20:09:43 +0000</pubDate>
      <link>https://dev.to/deixei/your-coding-agent-is-privileged-automation-a-workstation-baseline-for-claude-code-ddf</link>
      <guid>https://dev.to/deixei/your-coding-agent-is-privileged-automation-a-workstation-baseline-for-claude-code-ddf</guid>
      <description>&lt;p&gt;A coding agent reads the repository, runs shell commands with the developer's privileges, edits code and calls external tools. On the workstation, the least production-like place it runs, it is often left ungoverned. But distance from production does not remove the path to production.&lt;/p&gt;

&lt;p&gt;This is a condensed version of a baseline from the &lt;a href="https://www.deixei.com/ai-delivery-factory/" rel="noopener noreferrer"&gt;AI Delivery Factory&lt;/a&gt; reference, written for a regulated firm running Claude Code on managed Windows with WSL2, Linux workstations and hosted CI. It is vendor-specific on purpose.&lt;/p&gt;

&lt;h2&gt;
  
  
  The settings file is not the boundary
&lt;/h2&gt;

&lt;p&gt;Managed settings are a client-side control, not a security boundary. A local administrator can edit the managed source, and managed settings bind only Claude Code. Every high-risk context needs an OS or network control behind the policy file.&lt;/p&gt;

&lt;p&gt;Also: only managed settings and hooks &lt;strong&gt;enforce&lt;/strong&gt;. &lt;code&gt;CLAUDE.md&lt;/code&gt;, rules, skills and subagents &lt;strong&gt;steer&lt;/strong&gt; the model, so a control self-assessment should never claim a control on the strength of guidance.&lt;/p&gt;

&lt;h2&gt;
  
  
  Five layers
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Identity and tenancy:&lt;/strong&gt; SSO, a pinned login method and organisation, workload credentials for pipelines, no raw API keys on endpoints.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Managed policy:&lt;/strong&gt; one version-controlled document, delivered by device management and locked.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Runtime isolation:&lt;/strong&gt; the OS sandbox, plus dev containers for high-risk repositories.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Network and data:&lt;/strong&gt; egress allow-lists, tenant restrictions, secret placement, short local retention.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Monitoring and evidence:&lt;/strong&gt; OpenTelemetry export, hook events to the SIEM, and host process telemetry as the independent record.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Lock the merge
&lt;/h2&gt;

&lt;p&gt;Permission rules merge across scopes. A deny anywhere wins, but allows from user and project files combine with managed allows unless you lock them. The &lt;code&gt;allowManaged*Only&lt;/code&gt; flags close that separately for permission rules, hooks and MCP servers: locking the rules does not lock the MCP allow-list. Check it by sampling &lt;code&gt;/permissions&lt;/code&gt; on real workstations and reading the source column.&lt;/p&gt;

&lt;h2&gt;
  
  
  Denies are weaker than they look
&lt;/h2&gt;

&lt;p&gt;Evaluation runs deny, then ask, then allow. But rule matching is fragile: a deny on &lt;code&gt;curl&lt;/code&gt; does not catch &lt;code&gt;/usr/bin/curl&lt;/code&gt; or an &lt;code&gt;sh -c&lt;/code&gt; wrapper.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Use whole-command denies, not argument-constraining allows.&lt;/li&gt;
&lt;li&gt;Never allow-list environment runners such as &lt;code&gt;npx&lt;/code&gt; or &lt;code&gt;docker exec&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Put real parsing in a &lt;code&gt;PreToolUse&lt;/code&gt; command hook.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Denies don't cover arbitrary subprocesses or network access; the sandbox and the egress allow-list carry that load, and the agent should not be able to ask its way out of the sandbox.&lt;/p&gt;

&lt;h2&gt;
  
  
  Command hooks enforce, HTTP hooks carry evidence
&lt;/h2&gt;

&lt;p&gt;HTTP hooks fail open by default, so enforcement belongs in command hooks set to block on failure. Hooks and MCP servers run outside the sandbox, so review hook scripts as privileged code. Sending prompt content to the SIEM is a works-council and DPIA decision, not a default.&lt;/p&gt;

&lt;h2&gt;
  
  
  Know where the sandbox stops
&lt;/h2&gt;

&lt;p&gt;It covers shell commands and their children only; file tools, MCP and hooks run outside it. Native Windows and WSL1 have no sandbox, so make it fail rather than fall back silently. Default read access covers much of the machine, and a Docker socket is effectively host access.&lt;/p&gt;

&lt;h2&gt;
  
  
  Before you rely on it
&lt;/h2&gt;

&lt;p&gt;This baseline is a design proposal to pilot, not a certified configuration. Setting names, minimum versions and defaults change quickly, so confirm each one on the build you pin.&lt;/p&gt;




&lt;p&gt;The full baseline, with the threat model, identity routes, the seven-hook bundle and the regulated overlay: &lt;a href="https://www.deixei.com/ai-delivery-factory/build/coding-agent-workstation/" rel="noopener noreferrer"&gt;Coding-Agent Workstation Baseline (Claude Code)&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;This cross-post was condensed from the original with AI assistance.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>security</category>
      <category>devops</category>
      <category>claude</category>
    </item>
    <item>
      <title>Choose the least agency: how to pick the shape of an AI product</title>
      <dc:creator>DEIXEI</dc:creator>
      <pubDate>Sun, 15 Nov 2026 10:40:28 +0000</pubDate>
      <link>https://dev.to/deixei/choose-the-least-agency-how-to-pick-the-shape-of-an-ai-product-1ejg</link>
      <guid>https://dev.to/deixei/choose-the-least-agency-how-to-pick-the-shape-of-an-ai-product-1ejg</guid>
      <description>&lt;p&gt;Most AI products don't need an agent. The first design decision isn't which model or framework to use. It's &lt;strong&gt;who decides the next step: code or the model?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;This is a condensed version of a decision guide from the &lt;a href="https://www.deixei.com/ai-delivery-factory/" rel="noopener noreferrer"&gt;AI Delivery Factory&lt;/a&gt; reference, written for teams that build AI in regulated environments.&lt;/p&gt;

&lt;h2&gt;
  
  
  Five agency classes
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Class&lt;/th&gt;
&lt;th&gt;Shape&lt;/th&gt;
&lt;th&gt;Who decides the next step&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;Single call: one model call, no enterprise facts&lt;/td&gt;
&lt;td&gt;Code&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;td&gt;Grounded answer: retrieval plus one model call&lt;/td&gt;
&lt;td&gt;Code&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;2&lt;/td&gt;
&lt;td&gt;Workflow: a fixed sequence that may include model calls&lt;/td&gt;
&lt;td&gt;Code&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;3&lt;/td&gt;
&lt;td&gt;Bounded agent: one agent in a loop inside an envelope&lt;/td&gt;
&lt;td&gt;The model, within its envelope&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;4&lt;/td&gt;
&lt;td&gt;Multi-agent: several agents with separate owners and identities&lt;/td&gt;
&lt;td&gt;Models and code&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The rule is &lt;strong&gt;least agency&lt;/strong&gt;: start at the lowest class that meets the need, and climb only on evidence.&lt;/p&gt;

&lt;p&gt;The line between 2 and 3 is a control-flow test. If code selects the next step, it is a workflow, even when some steps call a model. Class 4 is defined by ownership, not by how many steps there are.&lt;/p&gt;

&lt;h2&gt;
  
  
  Eight questions before anyone builds
&lt;/h2&gt;

&lt;p&gt;Questions 1 to 5 set the agency class; 6 to 8 cover reuse, controls and hosting.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Is AI needed at all? Plain software ends many requests.&lt;/li&gt;
&lt;li&gt;Prediction or generation? Prediction belongs to classical machine learning.&lt;/li&gt;
&lt;li&gt;Does it need enterprise facts?&lt;/li&gt;
&lt;li&gt;Are the steps known in advance?&lt;/li&gt;
&lt;li&gt;Is it one coherent responsibility?&lt;/li&gt;
&lt;li&gt;Which catalogue assets cover each part of the application?&lt;/li&gt;
&lt;li&gt;Which policy pack applies?&lt;/li&gt;
&lt;li&gt;Which hosting route and model class do the data class and cost allow?&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Evidence for every step up
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;0 → 1&lt;/strong&gt; when answers need internal or current facts. Evidence: examples where the base model fails.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;1 → 2&lt;/strong&gt; when several distinct steps, tools or validations are needed. Evidence: the step list and single-call failure modes.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;2 → 3&lt;/strong&gt; when the next step can't be predicted and the paths are too many to code. Evidence: a path analysis and a cut-set answer.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;3 → 4&lt;/strong&gt; when tools or instructions conflict, or ownership, security or release needs are genuinely separate. Evidence: single-agent degradation results.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Four steps, four teams or a four-circle vendor diagram do not justify a multi-agent design.&lt;/p&gt;

&lt;h2&gt;
  
  
  Cap what it may do, separately
&lt;/h2&gt;

&lt;p&gt;Agency class says who picks the next step. How far the product may &lt;em&gt;act&lt;/em&gt; is a separate ceiling, set by the consequence of a wrong decision and the decision's complexity:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Consequence ↓ / Complexity →&lt;/th&gt;
&lt;th&gt;Low&lt;/th&gt;
&lt;th&gt;Medium&lt;/th&gt;
&lt;th&gt;High&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Low&lt;/td&gt;
&lt;td&gt;AI runs it&lt;/td&gt;
&lt;td&gt;AI runs it; human samples&lt;/td&gt;
&lt;td&gt;AI proposes; human decides&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Medium&lt;/td&gt;
&lt;td&gt;AI recommends; human approves&lt;/td&gt;
&lt;td&gt;AI recommends; human approves&lt;/td&gt;
&lt;td&gt;Human decides; AI assists&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;High&lt;/td&gt;
&lt;td&gt;Human decides; AI assists&lt;/td&gt;
&lt;td&gt;Human decides; AI assists&lt;/td&gt;
&lt;td&gt;Human decides; AI assists&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Decisions with legal or similarly significant effects default to the high row.&lt;/p&gt;

&lt;h2&gt;
  
  
  Cut-set review for agents
&lt;/h2&gt;

&lt;p&gt;For classes 3 and 4, ask one question: &lt;strong&gt;is any single control failure enough for the agent to act outside its mandate?&lt;/strong&gt; List every control between the agent and that event, test each alone and in pairs, and fix any single point of failure with a control that fails differently, not a better-worded prompt. The most common finding is a prompt instruction acting as the only barrier.&lt;/p&gt;

&lt;h2&gt;
  
  
  Write the decision down
&lt;/h2&gt;

&lt;p&gt;Every product gets a ten-field design record, committed before the first branch. The field that changes behaviour most: the chosen agency class &lt;strong&gt;plus every rejected lower class and why&lt;/strong&gt;. Listing what you rejected turns a receipt into a decision.&lt;/p&gt;




&lt;p&gt;The full guide, with the design record template, the cut-set method and a worked example, is here: &lt;a href="https://www.deixei.com/ai-delivery-factory/design/choose-ai-product-shape/" rel="noopener noreferrer"&gt;Choose the Shape of an AI Product&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;This cross-post was condensed from the original with AI assistance.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>architecture</category>
      <category>mlops</category>
      <category>devops</category>
    </item>
  </channel>
</rss>
