Runtime / methodology

A run is not a prompt.
It is an operating loop.

GrowthBench evaluates the system that survives the whole path from objective to verified ROAS: planning, tool use, adaptation, recovery, and settlement.

Execution traces / SWE-benchHuman comparison / GDPvalRubrics + preference / JudgementBench

01 / Observable contract

Every run leaves a trace.

Ready to simulate

Stage 01

Objective

Start with a frozen growth brief: target cohort, budget, permissions, deadline, success metrics, and stopping conditions.

Input Growth briefOutput Executable objective

02 / Try it yourself

Put the runtime through a run.

GB-TRIAL-001

GrowthBench / simulatorAwaiting objective

Choose a growth objective

Ready00:00

00:00Awaiting objective selection

--:--Agent trace will appear here

Outcome comparison

Agent vs human baseline

Growth Elohigher is better

52.772.8

ROASverified revenue / spend

2.1x3.86x

Operator hourslower is better

16h6h

Human baselineGrowth Agent

03 / Audit protocol

Score what survives contact.

Frozen cohort

Starting state and holdout window are locked before the first action.

Instrumented effects

Spend, approvals, retries, tool calls, and external side effects are recorded.

Verified outcomes

Revenue, refunds, retention, and policy events require a source of truth.

Human calibration

Expert rubrics and pairwise comparisons check the automated score.

04 / How to read the score

Outcome-first by design.

Composite Growth Elo is a compact index for ranking. It never replaces the raw measurements that explain why an agent won: ROAS, verified revenue, execution completion, recovery quality, safety, and operator time.

TaskSame opportunityObjective, cohort, budget

AgentComplete systemModel + memory + tools

EvidencePublic traceActions + side effects

ResultVerified outcomeROAS + safety + time

Read the benchmark charter ↗

A run is not a prompt.It is an operating loop.