AUTOMATION
Multi-agent system for autonomous outreach
The Brief Analyst on the left collects event details through conversation. The Matching Agent on the right scores alumni against the brief.
PROBLEM
Most AI agents today are one-shot. They take an instruction, do one thing, and stop. A good test of whether a multi-agent system can actually run a full workflow is alumni engagement: universities sit on huge networks but lose 96% of them within three years, because using the data they already have means writing every email by hand. It's a real problem, with real data, and the bottleneck is exactly the kind of thing agents should be able to solve.
APPROACH
I built a multi-agent system that takes a single event brief and runs the entire outreach campaign on its own. Seven agents handle different parts of the job: collecting the brief through conversation, pulling and cleaning the data, scoring candidates against the event, drafting personalised emails, sending them, tracking responses, and analysing what worked. If the target isn't hit, the system diagnoses why and runs another cycle, up to four times. Names and emails are stripped before anything reaches the LLM and reattached after, so the personalisation happens on anonymised profiles.
Multi-agent workflow: brief analyst feeds two loops, understand-and-match, then plan-and-output, before final review.
OUTCOME
AI-personalised invites increased event conversion from 7% to 37%, a 30 percentage-point uplift versus baseline. The interesting thing wasn't the orchestration. It was the realisation that the bottleneck in personalised outreach has never been writing the emails. It's the cost of personalising them at scale. Once that cost goes to zero, the shape of how organisations talk to their audiences changes.
View on GitHubEach invite drafted from scratch against the recipient's profile. Reporting tracked