From Open Prompts to Owned Workflows: What an Agent Harness Means for Designers Who Build
A new episode of "How I AI" quietly reframes how solo builders should think about AI agents — and the shift has real implications for designers running their own apps.
A recent episode of Lenny's How I AI series dropped three things worth paying attention to: a GPT-5.6 review, a look at how one solo builder runs a personal fleet of always-on local AI machines, and — most practically useful for designers who are starting to build — a clear explanation of what an agent harness actually is and why it beats typing into a general-purpose chat tool for any workflow you run more than once.
The thesis here is simple: the open chat window got you started, but a harness is what gets you consistent.
What Even Is an Agent Harness?
Let's kill the jargon first. An agent harness is just code wrapped around an AI agent to make it better at one specific job. That's it. Cursor (the AI coding editor many designers already use) is a harness. Claude Code is a harness. They're complex ones built by big teams — but the concept scales all the way down. As the episode explains, yours could be eight files and a terminal interface.
Think of a harness like a well-designed form versus a blank text field. A blank text field gives you infinite flexibility; a well-designed form gives you speed, consistency, and fewer mistakes. A harness does the same thing for an AI agent.
The episode's clearest example: the host built a custom harness using the Claude Agent SDK to automate bug triage at her company, ChatPRD. Every time a bug surfaces in Sentry (an error-tracking tool), the harness kicks off the same investigation sequence — gathering evidence, analysing root causes, and outputting a consistent bundle of files: a task log, a Sentry issue brief, relevant logs, a worker report, and an HTML summary. The engineering team gets a scannable record every time, without anyone writing it up manually.
Why This Matters If You're a Designer Building Your First App
If you've been vibe-coding — using tools like Cursor, Claude Code, or Bolt to build features through conversation — you've probably hit a moment where the AI goes slightly off-script. You forgot to say "don't touch the database" or "only investigate, don't write new code," and suddenly you're cleaning up something you didn't ask for.
A harness solves this by encoding your rules into the tool itself, not into every prompt. In the episode's Sentry harness, "investigate only, do not write code" isn't something you have to remember to say. It's a flag in the interface. You click it, paste a link, and the agent already knows its constraints.
For a designer building a product, this maps directly to real situations:
- You run the same user-feedback tagging process every week
- You review new content against a style guide on a schedule
- You check for broken links or accessibility issues every time you push an update
Any workflow that has defined steps, defined tools, and a consistent output is a candidate for a harness. The AI handles the judgment work in the middle; the harness handles the setup, the permissions, and the output format every time.
The Hybrid Model That Actually Works
The episode's conversation with solo builder Alex Finn adds another layer. Alex runs a small fleet of local AI machines — Mac Studios, a DGX Spark, an RTX 5090 — to keep agents running around the clock. The economics only make sense at that scale, so most designers don't need to go there yet.
But his underlying workflow logic is immediately transferable: use a cheaper, faster model for volume work, and a smarter frontier model for judgment calls.
His security scanning setup is the clearest illustration. A local model scans code every 20 minutes and dumps findings into a plain Markdown file. Claude Code checks that file once a day and decides what's actually worth fixing. The local model does the repetitive work cheaply; the powerful model does the discernment precisely.
You don't need a rack of Mac Studios to apply this. If you're using ChatGPT or Claude daily, you can already route: use a faster, cheaper model to do a first pass on something (summarising user feedback, flagging copy that's too long, spotting missing alt text), then bring your best model in only when a real decision needs to be made.
What to Actually Watch For
A few honest limits worth naming:
Building a harness still requires some code. The episode is candid that this is a step past pure prompting. If writing even a small amount of code feels out of reach right now, the more immediate takeaway is the mental model — start noticing which of your AI workflows you repeat, and treat those as harness candidates for later.
Tool connections add complexity. The Sentry harness works because it connects directly to Sentry and Linear (a project management tool). Wiring up integrations is where things can get fiddly, and the episode doesn't pretend otherwise.
General-purpose agents aren't going away. The episode is clear that just typing into Claude Code has produced real work — the shift is about recognising when a constrained, specialized tool will beat an open, powerful one. That's a when question, not an always question.
The Grounded Takeaway
The era of the open chat window as your primary building tool isn't over, but its limits are becoming visible. The designers and solo builders who are pulling ahead aren't necessarily using more powerful AI — they're using it more deliberately, with specific tools for specific jobs and consistent outputs they can actually hand off to a team or build on top of.
A harness is just a way of treating your AI workflow like a real product: something with defined inputs, predictable behaviour, and outputs you designed on purpose. You already think that way about interfaces. It's not a big leap to think that way about agents too.