Deep-Agents Harness & Flagship IB Products
Once hundreds of MCP tools piled up across the division, dumping them into one prompt destroyed tool-selection accuracy. We built a containerized skill execution engine and deep-agents harness with dynamic routing, workspace file access, layered memory, and human-in-the-loop checkpoints.
01 // THE CONSTRAINT
In private-side investment banking, approximate AI answers are useless if they cross an MNPI confidentiality boundary or hallucinate a single deal figure. On top of that, capabilities demoed in executive rooms often need to ship to production in days rather than quarters.
02 // THE ARCHITECTURE
I played a major role architecting and building the deep-agents harness that isolates specialized MCP servers, instructions, and domain agents inside composable skills. Rather than overloading context windows, the harness dynamically routes requests, injects personal preferences and organizational memory, reads workspace documents, and pauses on human-in-the-loop triggers when deal decisions require banker sign-off.
03 // THE OUTCOME
That single foundation powered three flagship workflows: client meeting preparation cutting senior banker prep from roughly 12 hours to 1.5 hours, a live deal collaboration workspace that helped close multiple billion-dollar-plus transactions, and automated pitch-deck generation adhering to strict corporate styling so analysts spend time on real research instead of aligning PowerPoint boxes.