virtual_rf

I spent 3 weeks vibe-coding apps and realized most of my time was going to debugging, not thinking. So I built an AI workforce manager, it sits above tools like Claude and Codex, runs an observe-steer-escalate loop across parallel workstreams, and persists what it learns across projects. The core idea. Tools like Claude and Codex are good workers, but nobody manages them. They take a task, execute it, and forget everything when the session ends. My system sits above them and runs an observe-steer-escalate loop. It watches what each worker produces across parallel workstreams, steers upcoming tasks based on what it learns, and escalates to me only when something is genuinely ambiguous. Crucially, it persists what it learns across projects. A discovery in Project A can inform a decision in Project B without me being the bottleneck.

But here's my dilemma. OpenAI is publishing about "harness engineering." Anthropic shipped Agent Teams. The platforms are pushing up the stack toward exactly this orchestration layer.

I'm torn between five options:

1. Keep it private – use it as my own power tool, demo selectively 2. Open source it – plant my flag, build community, get credit for the thinking 3. Source-available – visible but restricted commercial use 4. Commercialize a piece – extract one capability (like the observation protocol) into a focused product 5. Stop building, start learning – accept the platforms will solve this and become the best user of their tools instead

Full writeup with honest details on the architecture, gaps, and my financial reality here: https://rhysfisher.substack.com/p/i-built-an-ai-workforce-ma...

Here's what it feels like piloting the swarm via Telegram (https://x.com/virtual_rf/status/2025153021030592567?s=20)

Here's a small video expanding on some key ideas (https://x.com/virtual_rf/status/2025182435818799114?s=20)

What would you do?

I spent 3 weeks vibe-coding apps and realized most of my time was going to debugging, not thinking. So I built an AI workforce manager, it sits above tools like Claude and Codex, runs an observe-steer-escalate loop across parallel workstreams, and persists what it learns across projects.

The core idea. Tools like Claude and Codex are good workers, but nobody manages them. They take a task, execute it, and forget everything when the session ends. My system sits above them and runs an observe-steer-escalate loop. It watches what each worker produces across parallel workstreams, steers upcoming tasks based on what it learns, and escalates to me only when something is genuinely ambiguous. Crucially, it persists what it learns across projects. A discovery in Project A can inform a decision in Project B without me being the bottleneck.

But here's my dilemma. OpenAI is publishing about "harness engineering." Anthropic shipped Agent Teams. The platforms are pushing up the stack toward exactly this orchestration layer.

I'm torn between five options:

1. Keep it private – use it as my own power tool, demo selectively 2. Open source it – plant my flag, build community, get credit for the thinking 3. Source-available – visible but restricted commercial use 4. Commercialize a piece – extract one capability (like the observation protocol) into a focused product 5. Stop building, start learning – accept the platforms will solve this and become the best user of their tools instead

Full writeup with honest details on the architecture, gaps, and my financial reality here: https://rhysfisher.substack.com/p/i-built-an-ai-workforce-ma...

Here's what it feels like piloting the swarm via Telegram (https://x.com/virtual_rf/status/2025153021030592567?s=20)

Here's a small video expanding on some key ideas (https://x.com/virtual_rf/status/2025182435818799114?s=20)

What would you do?

Good question. It's worth a test.

Have you run the study in the real world? If you have, we also have the technology to be able to ingest studies and calibrate audiences so you get a benchmark and % lift.

If not, maybe there are some proxy data sets we can lean on to assess if the simulates responses are in the right ball park.

Totally fair. LLMs absolutely have limits. But calling it just regurgitation misses what makes them powerful. With the right setup, you can surface contradictions, edge cases, and patterns across millions of perspectives instantly. Figuring out the limits is part of using them well.

The point isn’t for AI to regurgitate the median view, it’s to give fast, diverse, and sometimes brutally honest feedback at scale. When prompted right, LLMs don’t just simulate consensus, they reveal edge cases, contradictions, and unfiltered perspectives humans might hide. Signal!

You mean trained to say they aren't right? Don't you think persona-of-thought prompting is more than just saying you're not an LLM...? (https://askrally.com/article/personas-of-thought)