Coding agent benchmark
Time to a working dev environment
A coding agent is only as fast as the environment it gets. We sent the same prompt to Aether, Replicas and Conductor Cloud on one shared repository with a Supabase stack, each product running its own startup hooks, and stopped the clock when the stack answered from inside the sandbox.
median of 5 fresh tasks, 17.3 s to 27.7 s
Replicas 31.9 s, and Conductor Cloud 73.5 s, medians of the same set of fresh tasks.
Prompt to a serving stack
Seconds from sending the prompt until the Supabase stack and the app answered a health probe. Lower is better.
- 19.4sAether: 19.4 s median, 17.3 s to 27.7 s over 5 runs
- 31.9sReplicas: 31.9 s median, 30.8 s to 32.5 s over 5 runs
- 73.5sConductor Cloud: 73.5 s median, 71.2 s to 80.1 s over 5 runs
- Aether17.3–27.7 s
- Replicas30.8–32.5 s
- Conductor Cloud71.2–80.1 s
- Products
- 3
- Fresh tasks
- 15
- Follow-ups
- 15
- CPU passes
- 33
Overview
One repository. Three products. One clock.
The repository holds the local Supabase stack, Postgres, PostgREST, GoTrue, Realtime, Storage and Studio, plus one Node app whose health endpoint answers only after it has queried Postgres through the stack. The same repository is configured in each product with that product's own startup hooks.
The agent's one shell call is the repository's probe script, which waits for the health endpoint and prints how long it waited. The clock stops when that result arrives, so the number includes every second the product spends before the agent can do useful work, and nothing the model does after.
Fresh tasks, follow-ups to a running task, and the CPU inside each sandbox are reported apart. Aether's fresh-task number is for a start that found a prepared seed.
Methodology
How the benchmark runs.
Every product ran Claude Code with Claude Sonnet 5.5 at low effort on the same commit of the shared repository, 5 fresh tasks per product with a follow-up message after each, through its public API from one machine. Timestamps are the client's own clock from the moment the request carrying the prompt was sent, polled four times a second. Runs took place on November 2, 2026: Aether and Replicas from 11:30 to 12:20 UTC, Conductor Cloud from 02:00 to 06:30 UTC.
Each product's hooks were set up the way its documentation describes and left at the product's defaults otherwise, except that Replicas ran at its 4 vCPU large size rather than its 2 vCPU default so the three sandboxes match on vCPUs: Aether bakes the dependencies, the pulled images and the running stack into a per-project seed; Replicas snapshots the images and dependencies in a warm hook and starts the stack in a start hook on every start; Conductor Cloud installs Docker in its organization install script and starts the daemon and the stack in a repository setup script on every new workspace.
The CPU row is the Phoronix Test Suite, version 10.8.4, running pts/node-web-tooling-1.0.1 in batch mode with its own run-count convergence inside each sandbox, reported as runs per second, every pass counted including the first after the sandbox was handed over.
The products do not sell identical machines; the table below names each sandbox's CPU, vCPUs and memory as read from inside it. The repository, with every product's hooks and the probe the agent runs, is public, so any row can be re-run, and this page carries each product's most recent run.
The sandboxes
CPU, cores and memory as read from inside each one.
| Product | CPU | vCPU | Memory | Startup hooks |
|---|---|---|---|---|
| Aether | AMD Ryzen 9 9950X | 4 | 8 GB | .aether/environment.json: setup, image prewarm and two terminals, baked into a seed |
| Replicas | Intel Xeon Platinum 8559C | 4 | 16 GB | replicas.json: a warm hook snapshots the images and dependencies, a start hook runs on every start |
| Conductor Cloud | Intel Xeon 2.9 GHz | 4 | 8 GB | organization install script and a repository setup script that runs on every new workspace |
Results
Fresh tasks, follow-ups, and the CPU underneath.
Fresh task: prompt to a serving stack
Seconds, median of the fresh tasks. Hover or focus a column for the range.
- 19.4sAether: 19.4 s median, 17.3 s to 27.7 s over 5 runs
- 31.9sReplicas: 31.9 s median, 30.8 s to 32.5 s over 5 runs
- 73.5sConductor Cloud: 73.5 s median, 71.2 s to 80.1 s over 5 runs
- Aether17.3–27.7 s
- Replicas30.8–32.5 s
- Conductor Cloud71.2–80.1 s
Follow-up on a running task
Seconds from sending a follow-up message until the probe answered again. Lower is better.
- 3sConductor Cloud: 3 s median, 2.9 s to 3.8 s over 5 runs
- 6.2sReplicas: 6.2 s median, 3.9 s to 6.5 s over 5 runs
- 16.7sAether: 16.7 s median, 15.6 s to 19.3 s over 5 runs
- Conductor Cloud2.9–3.8 s
- Replicas3.9–6.5 s
- Aether15.6–19.3 s
CPU inside each sandbox
Phoronix node-web-tooling, runs per second, every pass including the first. Higher is better.
- 25.7runs/sAether: 25.7 runs/s median, 14.2 to 29 over 15 passes
- 19.1runs/sReplicas: 19.1 runs/s median, 16.9 to 20 over 15 passes
- 15.3runs/sConductor Cloud: 15.3 runs/s median, 14.8 to 15.3 over 3 passes
- Aether14.2–29 runs/s
- Replicas16.9–20 runs/s
- Conductor Cloud14.8–15.3 runs/s
Your stack, running, in seconds.
Start a task on Aether and the dev servers are already up when the agent reads your prompt.