Debt Vulture

It started as a simple monitoring tool. Hedge fund investors in distressed debt were paying $30k-100k/year to incumbents like Octus and 9fin for bankruptcy filing alerts. We built a PACER scraper, a filing search dashboard, and email alerts — the same core product for $500-2k/month.

Then we added an AI research agent. It could search filings, read docket entries, pull SEC data, and answer questions about specific bankruptcies with citations. Standard Partners was using it to do in minutes what used to take analysts hours.

But the real question investors kept asking was: what happens next? Not "what has been filed" but "if I buy this tranche, what are the likely outcomes?" That's a simulation problem. So we started building agents that could represent the competing parties in a bankruptcy — creditors, debtors, sponsors, committees — and negotiate against each other.

The early version was bankruptcy-specific. But as we built the cognitive architecture — memory systems, an omniscient "God agent" referee, event-driven time, causality tracking — we realized the engine had nothing to do with bankruptcy. All the domain knowledge lived in the scenario description. The engine just provided cognition, memory, and consequences. Swap the scenario from a bankruptcy to a geopolitical crisis or a boardroom negotiation, and it works the same way.

So we made it domain-agnostic. The simulation engine is now a general-purpose multi-agent sandbox. Agents have generative-agents-style memory (adapted from Park et al. 2023) with importance scoring and reflections. Time is continuous, not tick-based — the engine jumps to the next interesting moment. Every action and consequence is tracked in a typed causality graph. You can branch simulations to explore counterfactuals. And the whole thing is accessible via a 27-tool MCP server so any AI agent can run simulations programmatically.

It ships with five historical bankruptcy presets — Serta Simmons, J.Crew, Hertz, Toys R Us — plus a quick demo scenario. But users can define anything. The frontend is a 7-panel lab IDE with live event streams, pacing metrics, agent inspection, and an interactive causality graph.

We surveyed the landscape in March 2026. Nobody has built a productized multi-agent simulation for distressed debt — or any negotiation domain — with real capital structure data and game-theoretic reasoning. The closest things are academic prototypes.