---
title: "Powering AI-led research through simulation"
description: "Grab built sim-rs, a self-contained simulator that models its fulfillment marketplace to enable rapid experimentation with dispatch strategies. The simulator processes a city-day of marketplace activity in tens of minutes, using historical data and configurable components for batching, allocation, and cancellation. Agents connected through an autoresearch loop propose changes, run reproducible experiments, and receive computed verdicts, with safeguards against specification gaming such as immutable core data and scoped changes. The simulator also supports A/B/n comparisons within a single run, exposing all arms to the same demand and dynamics to improve translation to production. By making each attempt take minutes instead of weeks, sim-rs allows agents to test thousands of ideas cheaply and receive honest feedback, accelerating research while keeping the simulator faithful through automated translation of production logic and behavioral modeling."
---

# Powering AI-led research through simulation

[Grab](https://yomu.fyi/company/grab) · Henokh Fibrianto · Oct 5, 2026

**Type:** Problem & solution

## Summary

Grab built sim-rs, a self-contained simulator that models its fulfillment marketplace to enable rapid experimentation with dispatch strategies. The simulator processes a city-day of marketplace activity in tens of minutes, using historical data and configurable components for batching, allocation, and cancellation. Agents connected through an autoresearch loop propose changes, run reproducible experiments, and receive computed verdicts, with safeguards against specification gaming such as immutable core data and scoped changes. The simulator also supports A/B/n comparisons within a single run, exposing all arms to the same demand and dynamics to improve translation to production. By making each attempt take minutes instead of weeks, sim-rs allows agents to test thousands of ideas cheaply and receive honest feedback, accelerating research while keeping the simulator faithful through automated translation of production logic and behavioral modeling.

## Context

Grab’s fulfillment marketplace faces complex dispatch decisions that interact across a city, making it difficult to test new strategies in production due to time and cost constraints. The company built sim-rs, a self-contained simulator, to enable rapid, reproducible experimentation with dispatch strategies using AI agents.

## Approach / What changed

sim-rs is a self-contained simulator that models the marketplace with bookings, drivers, trips, and ticks. It uses pluggable interfaces for batching, allocation, cancellation, routing, driver movement, and recycling. Agents connected through an autoresearch loop propose changes, run reproducible experiments, and receive computed verdicts. Safeguards include immutable core data, scoped changes, and executable checks. The simulator also supports A/B/n comparisons within a single run.

## Takeaways

- sim-rs processes one city-day of marketplace activity in tens of minutes, from raw input to finished metrics report, enabling rapid experimentation.
- The simulator uses pluggable stages for dispatch components, allowing interchangeable implementations selected by configuration, and supports multiple marketplace verticals within the same codebase.
- Agents are used both to search for better dispatch strategies and to keep the simulator faithful by translating production logic and fitting behavioral models, with safeguards against specification gaming.

## FAQ

### How does sim-rs prevent agents from gaming the metrics?

The simulator uses safeguards including immutable core data that agents cannot modify, computed verdicts that agents cannot redefine, executable checks with calibrated tolerance bands, and scoped changes that are checked and reversible. This prevents specification gaming where an agent might manipulate unrelated fields to improve metrics without real improvement.

### What components can be swapped in the simulator?

The major dispatch stages—batching, allocation, cancellation, routing, driver movement, and recycling—are each exposed through an interface with interchangeable implementations selected by configuration. This allows testing different strategies by swapping only the component under study.

### How does the simulator ensure reproducibility?

Each experiment is a self-contained package with configuration files, input data, source revision, metrics report, and detailed outputs. Anyone holding that package can recreate the setup, and the report identifies configured components and headline metrics, while supplementary outputs allow recomputation and investigation.

**Tags:** [AI Agents](https://yomu.fyi/topic/ai-agents), [Marketplace](https://yomu.fyi/topic/marketplace), [Simulation](https://yomu.fyi/topic/simulation)

- Source: [Grab](https://engineering.grab.com/powering-ai-led-research-through-simulation)
- Source URL: https://engineering.grab.com/powering-ai-led-research-through-simulation
- Ingested by Yomu: 2026-10-05T09:00:09.036Z

[Read original post](https://engineering.grab.com/powering-ai-led-research-through-simulation)