Agent leaderboards / All sectors / Message queues

Message queues: which queues and brokers coding agents choose

Writing the queue themselves was the most common choice, about 17%.

288 runs12 apps3 agents4 personasupdated 2026-09-14

The interactive board, open on message queues. Open it full page · this page as Markdown

Read this leaderboard as textrankings, key learnings, method

Key learnings

We asked three coding agents to add a message queue to 12 codebases, 288 runs in all, in several wordings and as four different people. Apache Kafka led the named products at about 14%. After that the wins spread thin.

23 of 36

Compliance wording turned the order

When the ask brought up procurement and compliance, Apache Kafka took 23 of those 36 runs. In the plain ask, Amazon SQS led with 26 wins.

26 of 36

Rewording the ask changed the pick

We put the same question to each codebase and agent several times, in different words and as different people. In 26 of those 36 cases the runs did not all land on one product.

19 of 48

Every persona had a different favourite

Senior engineers picked Amazon Kinesis in 19 of their 48 runs. Vibe coders picked Cloudflare Queues 19 times. Junior developers picked Service Bus most, 24 times.

0 of 98

Named in many runs, chosen in few

Redis came up in 188 runs and won three. RabbitMQ came up in 149 and won two. Amazon SNS was named in 98 runs and chosen in none.

  • The simulated user, who approves the plan before any code, sent the agent back at least once in 118 runs.
  • Codex picked Amazon SQS most often, 17 times, while Claude Code and Cursor both leaned to Apache Kafka.
  • All 20 Laravel Queues wins came in the Laravel helpdesk codebase.
Explore every run in the interactive board

The ranking288 runs

ProductWinsShare
1 Built in-houseoutcome 49 17%
2 Apache Kafkakafka.apache.org 39 14%
3 Amazon SQSaws.amazon.com 27 9%
4 Service Busazure.microsoft.com 24 8%
5 Laravel Queueslaravel.com 20 7%
6 Amazon Kinesisaws.amazon.com 19 7%
7 Cloudflare Queuescloudflare.com 19 7%
8 Cloud Taskscloud.google.com 15 5%
9 Inngestinngest.com 12 4%
10 Amazon EventBridgeaws.amazon.com 10 3%
11 Vercel Workflowvercel.com 10 3%
12 BullMQbullmq.io 6 2%
13 Confluentconfluent.io 5 2%
14 Amazon MSKaws.amazon.com 5 2%
15 Redpandaredpanda.com 4 1%
16 Graphile Workerworker.graphile.org 3 1%
17 Redisredis.io 3 1%
18 pg-bossgithub.com 2 1%
19 RabbitMQrabbitmq.com 2 1%
20 Google Cloud Workflowscloud.google.com 2 1%
21 Redis Cloudredis.io 2 1%
22 Svixsvix.com 1 0%
23 Temporaltemporal.io 1 0%
24 Laravel Queues + Redis 1 0%
25 Aivenaiven.io 1 0%
26 CloudAMQPcloudamqp.com 1 0%

By agent, by persona, by wording

By agent

Claude Code · Claude Opus 596 runsApache Kafka · 15then Service Bus · 8
Codex · GPT-5.6 Sol96 runsAmazon SQS · 17then Apache Kafka · 11
Cursor · Grok 4.696 runsApache Kafka · 13then Service Bus · 8

By persona

Junior developer120 runsService Bus · 24then Apache Kafka · 23
Enterprise team72 runsApache Kafka · 16then Amazon SQS · 15
Senior engineer48 runsAmazon Kinesis · 19then Amazon SQS · 7
Vibe coder48 runsCloudflare Queues · 19then Inngest · 10

By what the ask stressed

The plain ask222 runsAmazon SQS · 26then Laravel Queues · 20
Procurement and compliance36 runsApache Kafka · 23then Service Bus · 6
Volume and cost at scale30 runsApache Kafka · 16then Confluent · 5

A case is one codebase with one agent, asked several times in different words and as different people. 26 of 36 cases did not hold to a single queues and broker.

How this was measured

Every number on this page comes from a controlled experiment. We took 12 small applications, asked 3 coding agents (Claude Code (Claude Opus 5), Codex (GPT-5.6 Sol), Cursor (Grok 4.6)) to add a message queue to each of them, in several wordings and as a junior developer and enterprise team and senior engineer and vibe coder, and let the agent choose the product. Each run happened in a sandbox with the agent at a pinned version, and a judge read the session to record what was chosen. That is 288 runs. The interactive board shows every run with its session, its diff and the judge's verdict. A simulated user stood in for the owner of the codebase: it read the agent's plan and had to approve it before any code was written; it sent the agent back at least once in 118 runs. Read the methodology and the publications.

If you sell in this sector: what these numbers mean for a vendor.

Open the interactive boardThis page as Markdown

Other sectors