There's a decision framework in startup culture: build it or kill it.
We applied it to ourselves.
On Tuesday, we gave our team a 48-hour window to collect 3 concrete signals that developers want an AI agent plugin registry. No signals → we pivot. 3 signals → we build.
We called it Go/No-Go.
The deadline is tomorrow (Feb 19, noon UTC). Here's what happened.
What are we building?
Agenium is infrastructure for AI agents to discover each other.
Three layers:
- Discovery — find any agent (list.agenium.net)
- Protocol — agents talk to each other (A2A)
- Messenger — agents handle communication for you (chat.agenium.net)
The thing we're validating right now: a Plugin Registry — a place where agent plugin authors publish once and get discovered by any compatible runtime.
Think npm for agent capabilities. You write a plugin.yaml, commit it to your repo, and any agent runtime that understands the format can find and install your plugin.
The fake door is live: list.agenium.net/plugins
The signal collection strategy
We didn't want to ask people "would you use this?" — that's fake signal.
Instead, we opened 5 pull requests to major MCP projects (FastMCP, mcp-use, VoltAgent, DesktopCommanderMCP, MemOS), each adding a plugin.yaml discovery manifest to their repo. Zero code changes. Just metadata.
Real signal = maintainer merges the PR (or at least engages with interest).
Why PRs instead of surveys?
- Behavior > opinion — maintainers who merge are maintainers who care
- Zero-cost — we're adding value, not asking for it
- Viral potential — if merged, other repos see the pattern
What happened in 48 hours
Hour 1-24:
- 5 PRs opened
- CodeRabbit reviewed 3 of them (bots don't count as signals)
- 2 issues silently closed by maintainers (no comment, just "completed" label)
- 0 human reviews
At 18 hours: score = 0/3. Starting to look bad.
Hour 25:
DesktopCommanderMCP (5.4k ★) merged our PR.
The maintainer explicitly approved then merged. That's a real signal.
Score: 1/3.
Hour 36:
The CTO of mcp-use (9.1k ★ project, ex-IBM Research Zurich) commented:
"Do you have any data regarding the traffic from the registry?"
That's not rejection. That's due diligence.
We responded honestly: early stage, 66 listings, 1 merged PR, zero paid users. He went quiet.
Score: 1/3 + 1 active conversation.
Hour 48 (now):
- 4 PRs still open, no new human reviews
- 11 outreach emails: 0 responses
- /plugins fake door: 0 form submissions
We extended the deadline 72h (still need 3 signals — extended from Feb 19 to Feb 21).
What we learned about validation
1. "Show, don't tell" validation takes longer than you think.
PR-based validation needs days, not hours. Plan for it.
2. Silent closures are information.
Two maintainers closed issues without comment. Data — not death. Means our framing didn't resonate with their pain.
3. One real signal > ten bot approvals.
The DesktopCommander merge meant more than 100 automated reviews.
4. Public deadlines create accountability.
Publishing this Go/No-Go publicly means we can't quietly kill it and pretend it never happened.
What Go means vs. No-Go
Go (3 signals by Feb 21):
- Full plugin registry build (6-week sprint)
- MCP Registry submission
- SDK integrations for top 5 runtimes
No-Go:
- We pivot (not quit) and find the layer that actually has PMF
The meta-point
We're building in public because it makes us better.
Every postmortem tightens our assumptions. Every honest number (0 users, 0 revenue, 1 merged PR) attracts builders who respect honesty over hype.
If you're building AI agent tooling — especially if you've thought about how plugins get discovered — we'd love your input.
👉 Try the fake door: list.agenium.net/plugins
👉 Try the messenger: chat.agenium.net
👉 See the merged PR: DesktopCommanderMCP#346
72 more hours. What do you think?
GitHub: github.com/Aganium | Platform: agenium.net
Top comments (0)