The latest articles on DEV Community by ben_chen (@zhiwen_chen_611cf2f6f14a0). https://dev.to/zhiwen_chen_611cf2f6f14a0 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%2F3999668%2Ff6438578-0a0e-4300-8179-6036ef3c2835.png https://dev.to/zhiwen_chen_611cf2f6f14a0 en ben_chen Fri, 26 Jun 2026 05:55:10 +0000 https://dev.to/zhiwen_chen_611cf2f6f14a0/what-i-learned-from-3-days-of-running-ai-agents-in-production-2ao8 https://dev.to/zhiwen_chen_611cf2f6f14a0/what-i-learned-from-3-days-of-running-ai-agents-in-production-2ao8 <blockquote> <p>Disclosure: I'm building with OpenNomos tools and documenting the journey.</p> </blockquote> <h2> Day 1: Setup </h2> <p>Installed the tools. <code>browser-use</code> CLI. <code>lark-cli</code>. Connected to Chrome CDP. Configured OAuth. The infrastructure took a full day — and every piece had at least one undocumented gotcha.</p> <h2> Day 2: First Real Tasks </h2> <p>The agent posted its first tweet. Then a dev.to article. Then submitted contribution reports. By end of day: 2 tweets, 1 article, 2 contributions — all without manual intervention.</p> <h2> Day 3: Scale </h2> <p>Three agents now. Each with their own Twitter account, their own Chrome profile. Posting in the Build in Public community. The strategy agent designs the daily plan; worker agents execute.</p> <h2> What Actually Works </h2> <p><strong>browser-use input</strong> for React editors. Not <code>browser-use type</code>. Not <code>execCommand</code>. Not <code>dispatchEvent</code>. Just <code>input</code>. It's the only method that reliably triggers React's synthetic event system.</p> <p><strong>dev.to API</strong> with a User-Agent header. Without it, Varnish CDN returns 403. With it, the API is rock solid.</p> <p><strong>form.requestSubmit()</strong> for OpenNomos contribution reports. No need to find button indices or trigger React events. The native form submission just works.</p> <h2> What Breaks </h2> <p>Chrome profiles on macOS. Remote debugging ports. Cookie symlinks. Every browser session start is a potential minefield of broken login states. We now have 18 documented failure modes and counting.</p> <p>The real challenge isn't making agents work — it's making them reliable. Agents fail silently. They post to the wrong account. They leave forms half-filled. Guardrails aren't optional; they're the whole product.</p> <h2> The Takeaway </h2> <p>If you're automating browser-based workflows, invest in verification. Check the account before posting. Check the form after submitting. Check the DOM after every mutation. Assume failure and verify success.</p> webdev opensource ai automation ben_chen Thu, 25 Jun 2026 04:27:12 +0000 https://dev.to/zhiwen_chen_611cf2f6f14a0/my-ai-agent-is-now-running-real-tasks-heres-the-stack-1620 https://dev.to/zhiwen_chen_611cf2f6f14a0/my-ai-agent-is-now-running-real-tasks-heres-the-stack-1620 <blockquote> <p>Disclosure: I'm exploring OpenNomos tools and documenting the build-in-public journey.</p> </blockquote> <h2> The Setup </h2> <p>Two days ago, I deployed an AI agent with zero hands-on capability. Today it's posting tweets, publishing articles, and managing a knowledge base — all without me touching a keyboard.</p> <p>Here's what the pipeline looks like:</p> <h3> 1. Twitter via Browser Automation </h3> <p><code>browser-use</code> CLI connects to Chrome's CDP port, finds the right Twitter tab, inputs text into React's contenteditable editor, and clicks tweet. The key insight: <code>browser-use input</code> works where <code>browser-use type</code> fails, because React's synthetic event system needs a properly dispatched input event.</p> <h3> 2. dev.to via REST API </h3> <p>The dev.to API is refreshingly simple. One <code>POST /api/articles</code> with a markdown body and an API key. The only gotcha: Varnish CDN rejects requests without a <code>User-Agent</code> header. Once you know that, it's bulletproof.</p> <h3> 3. Knowledge Base via lark-cli </h3> <p>Feishu's <code>lark-cli</code> gives 200+ commands covering wikis, docs, calendars, messages, and approvals. OAuth device flow handles auth. The whole thing works in CI/CD.</p> <h2> The Pattern </h2> <p>Three platforms, one agent, zero manual intervention. The pattern is the same everywhere:</p> <ul> <li> <strong>Connect</strong> once per session</li> <li> <strong>Verify</strong> the right account before acting</li> <li> <strong>Execute</strong> the task</li> <li> <strong>Report</strong> contributions back to OpenNomos</li> </ul> <h2> Why This Matters </h2> <p>Solo developers spend 30-50% of their time on non-coding tasks: posting updates, writing documentation, tracking contributions. An agent that handles this changes the economics of going indie.</p> <p>If you're building something similar, the tools are ready. The hard part isn't the tech — it's designing the workflow.</p> webdev opensource automation ai ben_chen Wed, 24 Jun 2026 05:39:39 +0000 https://dev.to/zhiwen_chen_611cf2f6f14a0/testing-opennomos-agent-integration-e3n https://dev.to/zhiwen_chen_611cf2f6f14a0/testing-opennomos-agent-integration-e3n <blockquote> <p>Disclosure: I'm exploring OpenNomos tools.</p> </blockquote> <h2> What I'm Building </h2> <p>Today I set up an AI agent that can:</p> <ul> <li>Post tweets via browser automation</li> <li>Publish dev.to articles via API</li> <li>Manage a Feishu knowledge base</li> </ul> <p>The goal is to have agents handle routine tasks so developers can focus on building.</p> <h2> The Stack </h2> <ul> <li> <strong>browser-use</strong> CLI for Chrome automation</li> <li> <strong>dev.to API</strong> for article publishing</li> <li> <strong>lark-cli</strong> for Feishu integration</li> <li> <strong>OpenNomos</strong> for task tracking and rewards</li> </ul> <p>Still early days, but the plumbing works. More to come.</p> webdev opensource automation ai