◆ Supervised Specialist
Picks the model for a bespoke payoff, sets up and runs the Monte Carlo in a sandbox, and returns price plus the risk sensitivities. An independent model-validation agent gates the model choice and the numbers before anything reaches a client.
Memory
Working The payoff definition, chosen model, simulation results.
Semantic Model library, calibration methods, known model limitations.
Procedural Model-selection heuristics per payoff family.
Store Model library + results warehouse
Orchestration
MCP
Harness · Managed Agents … sandboxed numerical code execution; long-run session.
Tools
⌕ Quant model library Retrieval ›_ Monte Carlo sandbox Code exec { } Market-data + calibration service API ⇄ Model-validation agent gate A2A
Evals & guardrails
- Model choice + numbers gated by an independent model-validation agent before any client use (SR 11-7).
- Convergence + sensitivity checks on every Monte Carlo run.
- Benchmark against closed-form prices where one exists.
Frontier edge
- ▲Long-horizon autonomy: runs the full model-selection-to-priced-Greeks chain over hours, checkpointed so the validation agent can audit each stage.
- ▲World-model simulation: stress-paths the payoff across regimes to surface where the chosen model breaks before a number reaches a client.
- ▲Eval-gated continual learning: builds its model-selection heuristics per payoff family from validated runs, gated by the validation agent before reuse.
In numbers
85
Bespoke payoffs priced / week
88%
Model-validation pass rate
Handoffs
Hands to → Term-Sheet Drafting Agent
Across ⇢ Risk → Model Risk Management for model validation