The latest articles on DEV Community by Gan (@gan_e3fcc1c88e8). https://dev.to/gan_e3fcc1c88e8 https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3831434%2F4fc6215b-f2c5-45ab-8285-bffe107df6d5.jpg https://dev.to/gan_e3fcc1c88e8 en Gan Wed, 18 Mar 2026 16:08:34 +0000 https://dev.to/gan_e3fcc1c88e8/i-built-a-pipeline-orchestrator-to-improve-code-quality-beyond-single-shot-prompts-4fkl https://dev.to/gan_e3fcc1c88e8/i-built-a-pipeline-orchestrator-to-improve-code-quality-beyond-single-shot-prompts-4fkl <p>Despite improvements in coding agents, <strong>single-pass prompting rarely produces production-quality results</strong>. There’s still a lot of <strong>manual steering</strong> involved.</p> <p>My workflow lately has looked like this:</p> <ul> <li>Ask Claude to <strong>plan first (no code)</strong> </li> <li>Ask it to <strong>implement with constraints</strong> </li> <li>Ask it to <strong>review its own work</strong> </li> <li> <strong>Repeat reviews</strong> until no new issues are found</li> <li>Do a final human review before committing</li> </ul> <h2> The problem </h2> <p>Getting strong results from coding agents is already a <strong>multi-step process</strong>:</p> <ul> <li>Planning</li> <li>Implementation</li> <li>Iterative review</li> </ul> <p>But this workflow is <strong>not enforced</strong> — it lives in:</p> <ul> <li>Ad-hoc prompts</li> <li>Manual iteration</li> </ul> <p>As a result, it becomes:</p> <ul> <li><strong>Inconsistent</strong></li> <li><strong>Hard to reproduce</strong></li> <li><strong>Time-consuming to babysit</strong></li> </ul> <h2> The solution </h2> <p>I turned this into an <strong>automated pipeline</strong>.</p> <p>Instead of one agent doing everything, a <strong>manager coordinates multiple agents across stages</strong>:</p> <p><code>Plan → Implement → Verify → Review → Fix → Repeat</code></p> <p>Each step gets <strong>targeted prompts</strong>, not a generic instruction blob.</p> <p>For example:</p> <ul> <li> <strong>Planning →</strong> focuses on root-cause solutions, not workarounds</li> <li> <strong>Implementation →</strong> enforces full execution of the plan with well-documented code</li> <li> <strong>Verification →</strong> runs repo-specific commands (tests, linting, checks)</li> <li> <strong>Review →</strong> applies clear goals and severity calibration </li> </ul> <p>This is far more reliable than relying on a static <code>AGENTS.md</code> that may or may not be followed.</p> <h2> Workdocs and traceability </h2> <p>Every step writes to a persistent <strong>workdoc</strong>.</p> <ul> <li> <strong>Full context</strong> is carried across stages</li> <li>Each agent sees what previous steps did</li> <li>Humans can <strong>inspect the entire process</strong> </li> <li>Runtime evidence is logged for <strong>traceability</strong> </li> </ul> <h2> What this enables </h2> <p>Once the pipeline is enforced, a full system naturally emerges:</p> <ul> <li>Parallel execution via <strong>isolated git worktrees</strong> </li> <li> <strong>Dependency-aware task graphs</strong> (tasks run in waves)</li> <li>Automatic <strong>branch integration + conflict resolution</strong> </li> <li>Task-specific pipelines <strong>(feature, bug, refactor, etc.)</strong> </li> <li> <strong>Automatic task decomposition</strong> into subtasks</li> <li>Human <strong>approval gates</strong> at key stages</li> </ul> <h2> Result </h2> <p>Instead of managing prompts, you can <strong>just fire tasks at the system</strong>.</p> <p>And get:</p> <ul> <li><strong>Higher code quality</strong></li> <li><strong>Less babysitting</strong></li> <li><strong>More predictable outcomes</strong></li> <li><strong>Full visibility into what happened</strong></li> </ul> <p>It starts to feel less like prompting an assistant, and more like <strong>running a real development workflow</strong>.</p> <h2> Try it out </h2> <div class="highlight js-code-highlight"> <pre class="highlight shell"><code>pip <span class="nb">install </span>overdrive-ai <span class="nb">cd</span> /path/to/your/repo overdrive server </code></pre> </div> <h2> Repo </h2> <p>Open source (MIT):<br> <a href="https://github.com/Execution-Labs/overdrive" rel="noopener noreferrer">https://github.com/Execution-Labs/overdrive</a></p> <p>If you're exploring agent-based development workflows, would love your feedback.</p> <p><a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fjy4bqkrh7uihwl90cqqi.png" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fjy4bqkrh7uihwl90cqqi.png" alt=" " width="800" height="521"></a></p> webdev ai opensource machinelearning