OloLand’s new agentic architecture.
A Claude-style lead agent and subagent harness, but optimized for OloLand economics: Gemini 3.1 Pro by default, Claude for selective high-value turns, deterministic underwriting tools, fine-tuned risk SML, and an institutional learning loop.
Lead agent + subagents
Template YAML invokes a markdown lead agent. Agent() tool calls start isolated specialist loops for research, modeling, risk, forensic review, and writing.
Risk SML + RL
Fine-tuned risk extraction maps evidence to a structured risk taxonomy (300+ leaf-level checks). Corrections and deal outcomes create verifier rewards before model promotion.
IC-grade outcomes
A separate grader checks citations, evidence quality, assumption controls, and approval readiness before the artifact leaves shadow mode.
Reason. Plan. Gather. Act. Verify.
The lead agent walks a real Project Atlas QoE: extract financials, run Beneish, reconcile EBITDA against the source hierarchy, ship an IC-defensible answer.
Click the player and use spacebar / arrow keys to scrub the timeline.