The problem: a funnel that starts with homework
Moonlet operates validator and RPC infrastructure across blockchain networks. Its pipeline doesn't start with a phone call — it starts with homework. Every potential engagement is a project that must be found, screened, and understood: the technology, the team, the funding, the roadmap, the fit. Done properly, that's two to three hours per prospect. Done at pipeline scale, it's a full-time job that nobody was hired to do.
The result was familiar to anyone who has run a research-heavy funnel: shallow screening, stale notes, opportunities discovered weeks after the market did, and outreach written from memory rather than evidence.
What we built
We built Moonlet a single research workspace where projects, monitored sources, AI agents, research results, and activity history live together. No tab juggling, no copy-pasting between a scraper, a spreadsheet, and a doc. The workspace is the system of record for everything the team knows about an opportunity — and for how it came to know it.
The agents inside it do four jobs:
- Discover. An OpenAI-powered workflow reads collected web content and identifies projects worth a look — new networks, protocol launches, ecosystem announcements.
- Screen. A fast, inexpensive first pass scores each candidate against Moonlet's fit criteria, so the expensive work is spent only on prospects that deserve it.
- Investigate. A deeper research pass assembles the full briefing: analyses, summaries, and recommendations with the sources attached.
- Prepare. Research and outreach templates turn the briefing into a personalized contact strategy — positioning, talking points, and a draft first message.
Research that doesn't go stale
The part the team values most isn't the first briefing — it's the second one. The platform monitors external sources and triggers fresh research when something material changes: a funding round, a mainnet date, a team change. Prospect research stops being a snapshot that decays from the moment it's written and becomes a living file that updates itself.
The workflows themselves run on n8n, which means they're configuration, not code. The team can run an agent manually against a single prospect, wire it into the research pipeline, or adjust a workflow without waiting for an engineering release.
How a prospect flows through
Sources come in
Monitored feeds and collected web content land in the workspace.
Agents identify candidates
The discovery workflow extracts projects and de-duplicates against existing research.
Cheap screen, expensive dive
Screening scores fit; only qualified prospects get the full investigation.
Briefing and draft outreach
Analysis, recommendation, and a personalized draft — ready for review, never auto-sent.
Monitoring keeps it fresh
Material changes re-trigger research and surface in the activity history.
Outcomes
Figures from Moonlet's business-development workflow after adoption; screening throughput compared with the previous manual process.
What we learned
- Two-stage research is the economics of the whole system. A cheap screening pass in front of an expensive investigation pass is what makes agent research affordable at pipeline scale. Skipping the screen burns budget on prospects a human would have rejected in thirty seconds.
- Freshness triggers beat scheduled reports. Re-researching everything weekly produces noise. Re-researching when a monitored source changes produces signal — and the activity history shows why the update happened.
- Templates are guardrails, not constraints. Giving agents an explicit research template and outreach structure improved output quality far more than better prompts did. Structure is what makes agent output reviewable.
- Drafts, not sends. The system prepares; people decide. Keeping the human on the send button costs seconds and preserves the thing outreach depends on — judgment.
Where this pattern applies
Nothing in this architecture is specific to blockchain. Any research-heavy funnel has the same shape: partner scouting, market monitoring, vendor due diligence, grant screening. If your team spends hours assembling context before every conversation, the same pattern — monitored sources, two-stage agent research, human-approved output — applies. Inside a Microsoft environment, the natural home for it is Teams and Dynamics 365, with the same agents doing the homework.
The one-line summary: we didn't automate the relationship — we automated the two hours of homework before it, and made sure every claim in the briefing carries its source.