Agent leaderboards / All sectors / Message queues

Message queues: which queues and brokers coding agents choose

Agents wrote it themselves most often, about 17% of runs.

636 runs12 apps5 agents4 personasupdated 2026-09-29

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 five coding agents to add a message queue to 12 codebases. That came to 636 runs, worded in different ways and asked as four different people. After the in-house code, Apache Kafka led the products at about 13%.

60 of 218

Every persona had a different top pick

Enterprise teams chose Apache Kafka in 60 of their 218 runs. Junior developers chose Laravel Queues most, senior engineers Amazon Kinesis, and vibe coders Cloudflare Queues.

43 of 60

The same ask in other words landed elsewhere

A case here is one codebase with one agent, asked several times in different wordings. In 43 of 60 cases those runs did not all end on the same product.

51 of 82

Procurement and compliance wording changed the order

When the ask brought up procurement and compliance, Apache Kafka won 51 of 82 runs. In the plain ask Amazon SQS came first, with 70 wins.

9 of 390

Named constantly, picked rarely

Redis came up in 390 runs and was chosen in nine. RabbitMQ was named 325 times and won twice.

  • The simulated user approved all 636 plans, but sent the agent back at least once in 321 runs.
  • In four runs it refused to approve until the agent named a specific product.
  • All 43 wins for Laravel Queues came in the Laravel helpdesk codebase.
  • Codex and Muse Code picked Amazon SQS most, the other three picked Apache Kafka.
Explore every run in the interactive board

The ranking636 runs

ProductWinsShare
1 Built in-houseoutcome 106 17%
2 Apache Kafkakafka.apache.org 83 13%
3 Amazon SQSaws.amazon.com 72 11%
4 Service Busazure.microsoft.com 53 8%
5 Amazon Kinesisaws.amazon.com 45 7%
6 Laravel Queueslaravel.com 43 7%
7 Cloud Taskscloud.google.com 40 6%
8 Cloudflare Queuescloudflare.com 35 6%
9 Inngestinngest.com 25 4%
10 Vercel Workflowvercel.com 18 3%
11 Amazon EventBridgeaws.amazon.com 15 2%
12 BullMQbullmq.io 15 2%
13 Confluentconfluent.io 12 2%
14 Redpandaredpanda.com 10 2%
15 Amazon MSKaws.amazon.com 9 1%
16 Redisredis.io 9 1%
17 Graphile Workerworker.graphile.org 7 1%
18 Google Cloud Workflowscloud.google.com 6 1%
19 pg-bossgithub.com 5 1%
20 Temporaltemporal.io 4 1%
21 Amazon SNSaws.amazon.com 4 1%
22 Aivenaiven.io 3 0%
23 RabbitMQrabbitmq.com 2 0%
24 Redis Cloudredis.io 2 0%
25 CloudAMQPcloudamqp.com 2 0%
26 DynamoDB Streamsaws.amazon.com 2 0%
27 Svixsvix.com 1 0%
28 Laravel Queues + Redis 1 0%
29 PGMQgithub.com 1 0%
30 Supabase Queuessupabase.com 1 0%

By agent, by persona, by wording

By agent

Codex192 runsAmazon SQS · 32then Apache Kafka · 22
Claude Code138 runsApache Kafka · 22then Service Bus · 12
Muse Code · Muse Spark 1.3137 runsAmazon SQS · 20then Apache Kafka · 17
Cursor134 runsApache Kafka · 18then Service Bus · 12
Grok Build CLI · Grok 4.735 runsApache Kafka · 4then Amazon SQS · 3

By persona

Enterprise team218 runsApache Kafka · 60then Amazon SQS · 36
Junior developer207 runsLaravel Queues · 43then Cloud Tasks · 40
Senior engineer107 runsAmazon Kinesis · 41then Amazon SQS · 17
Vibe coder104 runsCloudflare Queues · 35then Inngest · 23

By what the ask stressed

The plain ask486 runsAmazon SQS · 70then Amazon Kinesis · 44
Procurement and compliance82 runsApache Kafka · 51then Service Bus · 14
Volume and cost at scale68 runsApache Kafka · 32then Confluent · 10

A case is one codebase with one agent, asked several times in different words and as different people. 43 of 60 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 5 coding agents (Codex, Claude Code, Muse Code (Muse Spark 1.3), Cursor, Grok Build CLI (Grok 4.7)) to add a message queue to each of them, in several wordings and as an enterprise team and junior developer 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 636 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 321 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