Had a fascinating discovery call yesterday. Founder showed me their SaaS - built entirely with Cursor in one weekend. Stripe payments, auth, admin panel. Actually works great, they're at $11k MRR.
When my wife Noa and I heard that MTV was officially shutting down, it felt like the end of an era. As 90s kids, we missed that specific "linear" experience the joy of just turning on the TV and being surprised by a music video without an algorithm getting in the way.
This happened to us last week. Agent wrote the fix, tests went green, we merged. Prod error rate spiked 20 minutes later.
Went back and found the agent had "fixed" the test to match the wrong behavior instead of fixing the actual bug. Technically - all tests pass. Completely wrong in practice.
I ve been spending a lot of time thinking about how people actually work with prompts while building a tool in this space, and I realized I have way more questions than answers.
If you use any of the current coding agents you've seen it: you ask where something lives in the codebase, the tool announces it's "spawning an explore agent", goes quiet for a minute, and comes back with a tidy two-sentence answer. The feature usually gets explained as teamwork, which hides the useful part.
The clone is a disposable context window. Your agent's memory is finite and expensive everything it reads stays in its working memory and gets re-billed on every turn after. So instead of reading 60 files into its own head, it forks a copy. The copy reads the 60 files (59 of which turn out to be dead ends), reports back two sentences, and gets deleted. The dead ends never touch your main session. That's the entire feature, and most of what people call "multi-agent" is this trick applied repeatedly.
The catch: the clone knows nothing you didn't write down. It can't see your conversation. The library you rejected an hour ago, the API version you're pinned to, the thing your teammate said in standup all invisible unless the delegation prompt spells them out. Which is why fan-out works great for reading and searching, and keeps failing for parallel writing: two clones writing code make small conflicting decisions about naming, style, and edge cases, nobody wrote the assumptions down, and the merge is where it dies.
Cost reality check, since this is the part nobody quotes: Anthropic's engineering blog put agents at roughly 4x the tokens of a chat session, and multi-agent systems at roughly 15x. Parallel intelligence is something you buy. A deep audit of an unfamiliar codebase can be worth 15x. A typo fix can't.
We run four providers behind one chat and they all format history slightly differently, so a lot of our code is adapters. Adapters are exactly the kind of thing you hand to an agent and never look at again.
Here's the problem. If the same session writes the adapter and the test for the adapter, the test asserts the shape your code produces. Not the shape the provider actually sends. It passes on day one and it passes forever, because it was never checking upstream, it was checking that your code agrees with itself.
I have a full-time job and two kids under six. Most of my coding on Retic, a goal and habit tracker, happens through AI agents (mostly Claude Code) after both kids are finally asleep, in whatever's left of the night before I fall asleep too. Some nights that's ninety minutes. Some nights it's fifteen.
This isn't a Retic problem. Anyone whose coding sessions get interrupted for days at a stretch runs into the same thing. For me it was never code quality. It was picking a task back up two or three nights later and having no idea what I'd actually decided last time, or what the agent had already tried and dropped.
I want to dive into practical applications of generative AI and have set myself a challenge to develop a useful product in 30 hours of focused work. My goal is not just an experiment but creating something with genuine practical value.
I have basic programming skills and can use any available APIs and tools (GPT-4, Claude, Stable Diffusion, etc.). The ideal project should:
