Show HN: AI agent that runs real browser workflows
ghostd.ioI’ve been experimenting with letting an AI agent execute full workflows in a browser.
In this demo I gave it my CV and asked it to find matching jobs. It scans my inbox, opens the listings, extracts the details and builds a Google Sheet automatically.
Cool demo. The tricky bit with browser workflow agents is figuring out which workflows to automate in the first place. Most people don't even realize they're doing the same thing over and over - they just do it.
I've been building MemoryLane (https://github.com/deusXmachina-dev/memorylane) which comes at this from the other side - it records screen activity, spots repeated patterns with AI, and then tells you "hey you keep doing this, want to automate it?" Works as an MCP plugin for Claude/Cursor.
Feels like pattern detection (finding what to automate) + browser agents like yours (actually doing the automation) is the right combo. Are you thinking about the discovery side at all, or mostly focused on execution?
Interesting. Part of why I built this was to avoid screen capture as the control layer. Once you’re taking screenshots, guessing what to click, moving the mouse, and repeating, it gets slow and brittle fast. Here the workflow is just described in text, executed in the browser, and saved for reuse.
The CV-to-job-search demo is a good showcase. For multi-step workflows like this, one issue I've run into is that the agent instructions degrade across steps. The initial task description is clear, but by step 5 the model is interpolating intent from earlier context rather than following explicit instructions.
Structuring the task prompt into named blocks (objective, constraints, expected output format per step) before the workflow starts makes each step much more reliable. The agent has less to infer.
Built github.com/Nyrok/flompt to help with this, a visual builder that decomposes instructions into semantic blocks and compiles to Claude-optimized XML. Useful for defining the "task shape" before handing it to an agent.
Yeah, instruction drift is a real problem in long agent chains. In this case the workflow gets decomposed into steps up front and each step is executed by a separate sub-agent.
So the model isn’t carrying the whole instruction chain across multiple steps, it’s just solving the current task. Similar pattern to what tools like Codex CLI or Claude Code do.
Interesting demo, how are you thinking about prompt injection and security with web agents? Ive been facing this as well.
Prompt injection is the same problem all agents face, ChatGpt Atlas, claude cowork, openclaw, all of them. It's a known unsolved problem across the industry.
I mitigate it by giving the agent a fixed action set (no scripts, no direct API calls), and breaking tasks into focused subtasks so no single agent has broad scope. The LLM prioritises its own instructions over page content, but if someone managed to hijack it, the agent can interact with authenticated sessions. Everything's visible in real time though, and all actions are logged, so you can see exactly what it's doing and kill it.
Practically speaking, I use it similar to how people use Zapier or n8n, you set up specific workflows and make sure you're only pointing it at sites you trust. If you're sending it to random unknown websites then yeah, there's more risk.
But even then, an attacker would need to know what apps you're authenticated with and what data the agent has access to. The chances of something actually happening are pretty low, but the risk is there. No one's fully solved this yet.
I was looking for a similar produc/project the other day. Alas my need is a Linux native version. You may want to consider it as Mac seems to be overserved by the agent harness supply while Linux is the opposite
linux and windows support is on the way, i’ve designed it in a decoupled way, so should be straight forward.
Just need to see if people find this version useful