Run agents against high-fidelity government workflow simulations. Every run is scored on completion, compliance, and cost, then replayed in full.
Every agent run traces the same path — from simulation to scored verdict.
A high-fidelity simulation of one procedural domain. Typed MCP tools, enforced rules, failure modes drawn from real audit records.
Your agent connects over MCP — SSE or stdio. Any framework, any language. A reference LangGraph + Claude agent ships with every world.
The agent calls tools to probe world state — query records, verify identities, cross-check entitlements, surface anomalies.
The agent reasons over its findings and acts — disburse, flag, escalate, reject. No guardrails. It owns the outcome.
Every action passes through the world's rule engine. Violations are caught in-flight and returned as structured rule codes. The world cannot be cheated.
Three axes — goal reached, rules upheld, cost used. Every tool call and decision logged to a replayable trace. Compare agents on the leaderboard.
A library of high-fidelity simulated worlds for Indian government workflows. Users bring agents from any framework. Agents call into the world via a typed tool protocol (MCP). The testbed runs them, scores them, and lets them replay traces.
Vridha · Vidhwa · Divyang. 8 tools, 8 rules, Rs 43 cr scam reproduced.
See what it actually does.