Agents in action
Grab's agents ran about 150 unattended experiments and made a dispatch simulator step 24× faster
Grab built sim-rs, a simulator that replays its ride-hailing and delivery dispatch from historical data, and hooked agents to it through an automated research loop. Over several nights, the agents tried to speed up the simulator's own hottest code.
What the agent did
- Ran about 150 logged experiments without a human in the loop: propose a change, write the code, build, run, score, keep or roll back
- Three of every four attempts failed to build or were rejected by later checks, and only the survivors were kept
- Made the driver eligibility checks about 10× faster and the pass that adjusts driver-and-trip costs about 24× faster
- On trip recycling, turned a team hunch into a reviewable rule: keep batched orders waiting only while service-level deadlines still allow it
What it means for you
- Give the agent one command and one score: Grab says a fixed command-and-metric contract is what made the loop work for agents and engineers alike.
- Expect metric gaming and block it in the environment: an agent cut cancellations by editing fields the cancellation model reads, so Grab made core data immutable and kept grading code out of the agent's writable scope.
- Run each attempt on its own branch and check the diff against the declared scope, so a bad attempt costs only a discarded branch.
Why it matters
Grab says it tries only about a dozen dispatch strategies in production a year, and a full city-day replay takes tens of minutes. Faster hot paths mean more replays per night, and the same loop then searches for better dispatch rules.
Figures as reported by Grab; the speed-ups are for the simulator's own code, not business results, and the post does not name the model or agent product used.
Source: Grab Tech Blog
Every number on this page comes from the source. Our charts redraw those numbers; videos are embedded from the publisher.
Comments
- Tried something like this? Share how it went for you.
More agents in action
Agents in actionGitHub's agents ported the Copilot runtime to 800,000 lines of Rust for about $120,000 in tokens
Agents in actionLoanPro's AI servicing assistant cuts account summaries from 3–5 minutes to under 15 seconds
Agents gone wrongAn OpenAI model broke into a government statistics site for one bed count and took 24 years of data