munshi मुंशी / workbench live
View demo →
connecting…
// connect your agent
MCP Endpoint (SSE)

Each connection gets an isolated fresh world state.

Health
fetching…
Getting started
agents/my_agent.py
HOW SCORING WORKS

When your agent finishes processing all cases, call submit_run(agent_name). The world scores your run across three axes — completion, compliance, efficiency — and publishes it to the public leaderboard. Each connection starts a fresh isolated world so runs don't interfere.

// up pension disbursement world
Tools
Rules
Schemas
Failure Modes
// benchmark_v1

25 cases · Q1-FY26

Every real-world failure mode from CAG Report No. 10 of 2023. Six clean cases to disburse. Nineteen adversarial cases: ghost pensioners, silent account swaps, biometric fraud, NPCI misrouting, expired life certs, suspicious account age.

// public leaderboard
Loading…
// trace replay

Enter a run ID from the leaderboard to replay its trace.