Flue, Eve, and the One Agent Idea That Didn't Feel Fake

Flue, Eve, and the One Agent Idea That Didn't Feel Fake

I made a tool that checks Trello. It works fine. Instead of nice Kanban cards, I now get a mediocre text reply from a chat app. This is not a better experience.

I made another that fetches the weather. Cmd+Space, type "weather," done. I did not need to ask an AI to do this.

This is basically the story of trying to figure out what to actually build with an agent framework, and it didn't take long to admit that most of my ideas were just worse versions of things that already worked.

The tech bro at 1000%

You've seen the tech bro on socials with the fully agentic OS: a dozen sub-agents running his calendar, his inbox, his life. Anyone else think this is sad as fuck? I used to think the bullet journal crowd was over the top. They seem normal now. What happened to just being human, having the basic life skills to manage your own schedule?

I think that's half my problem too. I get an idea, and my first instinct is just do it yourself, man, for ffs. So the Trello checker and the weather bot both exist purely because I wanted to test tool calls, not because anyone needed them. They're getting ditched.

Two frameworks, on purpose

If you don't know Flue or Eve, they're agentic frameworks: bundled tooling for building your own agents. That's a big deal for me. I've got a pile of ideas I want to build with AI, and I could not wire this up myself. The complexity, and how fast this space moves, would break me.

I'm using both, deliberately. The chat app I'm building for Zero Saints runs on Eve, on Vercel. The marketing bot on our site is already Eve, so it made sense to stay in that world, and the dev experience is straightforward. My personal assistant, Aury, runs on Flue, on Cloudflare.

Why split them? Because I think this is just what AI looks like right now, moving in what feels like 4D space, genuinely exhausting to keep up with. Staying in the loop means context-switching between stacks I'd otherwise just pick one of. So be it.

The actual idea

Here's what's real: a collaborative workspace for me and my partner to run our business through.

Right now our shared files live on Google Drive, and most of them are AI-authored these days. The flow looks like this: we talk through an idea (breakfast, a drive, going for coffee) → she takes it to her AI of choice → it produces a markdown SOP or summary → uploads to Drive → I download it, take it to my AI, we work on it → I re-upload, and now I'm either overwriting the original or making a duplicate and renaming it, which gets messy fast.

The real failure mode: our agents end up referencing stale files. She might have the original on her MacBook, I've got the edited version, and she hasn't pulled it down. We're out of sync, and it's a chore to notice, let alone fix.

What I actually want is git. One chief agent, Aury (named after our bossy cat), sitting in the middle, organising files, comparing diffs, helping us actually converge on one version of a document instead of quietly diverging. Even if we go off and use Claude or ChatGPT separately, the moment either of us uploads back into our own filesystem, Aury reconciles it. I haven't nailed the exact shape of this yet, how much Aury resolves versus just flags for us, but that's the part of this I'm actually excited to build, rather than the part I'm building because I can.

(Yes, Syncthing exists. That would be too simple.)

What lives in it

A few things I want Aury to actually do:

• A shared copywriting skill. Same tone and language whether I'm writing a Zero Saints blog post or my partner's writing a social caption. This is exactly where voice drift happens without shared context.
• Last30Days, shared. The whole point of it goes stale without joint context: daily fresh content for our niche that either of us can dig into.
• Error triage. I get a stream of noise from Vercel, Sentry, and friends. I don't read most of it. I'd rather Aury summarise it and tell me when something's actually wrong.

Why not just pay for it

Claude and ChatGPT have had some version of shared project context since 2024. We could get that collaboration on a Team plan for something like $43/each a month (worth double-checking current pricing before you take my word for it). It'd probably work fine. I'm building my own anyway; it's more rewarding, and it does exactly what I want instead of what the plan tier allows.

The cats have jobs now

Spinning up sub-agents is stupidly easy in both frameworks, so (sticking with the household theme) the cats have roles:

• Cato: a bit of a ponce, academic, so researcher.
• Macro: territorial, so security agent, watching Sentry and Snyk.
• Luna: missing her tail, finds climbing hard, so accessibility.
• Peach: a diva, obviously marketing.
• Gisela: territorial, agile, role TBD.
• Carlos: our big old Maine Coon, senior dev.

And Aury's the chief, coordinating all of them. Before anyone panics at the cat count: they're not all ours, they're spread across the family.

Why bother

Not everything has to be about making money. This is the reminder I keep needing: half of what makes this fun is that it doesn't have to justify itself. I'm learning a lot on both Flue and Eve just by building something dumb and specific instead of something impressive and generic, and that's reason enough on its own. More on both as I get further in.