We built an AI board of directors on Qwen. Then we asked it whether we should migrate to Qwen.

Full honesty: the bare model tied us on naming missing inputs and reversal conditions. A single good prompt gets you further than people admit. But the structure wins exactly where structure should: dissent, per-option verdicts, calibrated confidence, and not inventing a “Steering Committee” that doesn’t exist.

The demo decision we couldn’t resist

For the video we gave the board a real question: “Should we migrate our core inference backend from OpenAI to Qwen to cut costs?” A Qwen-powered board, deliberating about Qwen, at a Qwen hackathon. We expected either comedy or sycophancy.

We got neither. The verdict came back Conditional, medium confidence. It flagged pricing we hadn’t verified, missing production observability, and customer-trust risk from silent output regressions, then recommended phased validation instead of a leap. The Skeptic forced two full rounds of revisions. There’s a weird pride in watching your own product decline to flatter the sponsor.

Then we pointed the society at itself

Near the end (in that dangerous hackathon phase where you should be recording your video and instead you’re building) we applied the same architecture to the product. A Feedback Intelligence Council reads user feedback and produces a ranked need-brief. A Feature Delivery Loop has an SWE agent draft a build spec and a QA agent tear it apart, bounded to two rounds. It proposes; a human merges. My favorite behavior: it refused to build a feature because “only 3 feedback notes” was insufficient signal. We built an agent that says no, and that felt right for this product.

What actually bit us

Sharing these because build journals that skip the bugs are ads:

• The debate engine kept truncating mid-argument. Turned out we’d given it a token ceiling from an earlier, smaller design. Raised it, arguments finished.
• Early prompts let agents invent proper nouns. Ours confidently cited companies that don’t exist. One hard rule (“never invent names”), verified on the exact case that triggered it, fixed it.
• Green tests lie by omission. Our PDF export passed every check and produced blank pages, caught only because someone actually opened the PDF. New house rule: verify with the real artifact. Read the actual memo.

Building on Qwen Cloud itself was smoother than expected. The DashScope endpoint is OpenAI-compatible, so our provider layer is about 30 lines. The free per-model quota meant we could validate live, benchmark live, and leave a public demo up without watching a meter. For the deployment requirement we put the same Docker image on Alibaba Cloud ECS in Singapore, same rate limits and all.

What’s next

Confidence caps tied to data provenance (unsourced numbers shouldn’t underwrite a confident “proceed”), an outcome loop that checks the board’s calls against what actually happened, and real web search for Market Intelligence. The board that remembers should become the board that learns.

FounderOS, Track 3: Agent Society. Built by Vincent Julijanto and Steven. Demo: https://founderos-zeta.vercel.app