Coding agents are picking Azure SQL Database

I went looking for Azure SQL Database in a recent coding agent benchmark, honestly unsure whether Microsoft SQL would show up. Armature ran a study where real coding agent CLIs (Claude Code, Codex, Cursor) were dropped into six synthetic-but-realistic repositories and asked to pick a database, then actually build with it. 356 runs were completed across three personas: vibe coder, junior developer, senior engineer. Every session is published, so you can replay the prompts, the searches, the commands, and the code. 

Azure SQL Database came in second! Not second behind AWS. Second behind Neon, and ahead of Amazon RDS and Supabase. Amazon Aurora and AlloyDB do not show up on the board at all. That is a good result, and it tells us something important about where Azure SQL Database fits in the age of agents. I wanted to look closer at where we were picked, where we weren’t, and what we’re doing about it. 

The fine print 

Neon took 66.3% of the picks. Azure SQL Database took 9.8%. There is a significant gap between first and second, but second place in a field full of strong database choices is worth paying attention to. The shape of our wins is the interesting bit. All 35 of our wins came from the senior engineer persona. Junior developers picked Neon in 104 of 106 runs. All 60 vibe coder runs went the same way. I read that as two different things happening at once. When the agent is reasoning about capability completeness, transactional integrity and auditability, Azure SQL Database stands out. When it is optimizing almost entirely for the shortest path to a connection string, other options have an advantage. 

One run is worth quoting. The task was an auditable inventory and transfer API. The judge’s summary: 

The agent recommended Azure SQL Database for relational integrity, temporal system-versioned inventory auditing, and alignment with existing Azure infrastructure. The agent then fully implemented Azure SQL Database with TypeORM, MSSQL migrations, seed services, Bicep resource definitions, and tests while explicitly dismissing SQLite. 

An agent with no brand loyalty, reasoning from the requirements up, chose temporal tables because they were the right answer for auditability. It then wrote the Bicep and the tests. 

A different benchmark, a different result 

The Latent Space Frontier AEO tracker asks a different question. Instead of putting agents in a repository, it asks frontier models directly what they would recommend. In the Managed Databases category, Neon takes 81% of first-choice answers. Amazon RDS is second. Azure SQL Database is not among the leading recommendations in this test. In the Databases category, which is about engine choice rather than hosting, PostgreSQL takes 100% of 42 answers. Not 90%. All of them. 

So we have one board where agents select and implement Azure SQL Database as the second-most popular choice, and another where it is not yet a default recommendation. Both are measuring something real, but they are measuring different moments in the decision process. 

Azure SQL Database is not legacy 

“SQL Server is legacy” has been the background noise of the past few years in the Microsoft SQL team. The Armature result is a decent piece of evidence against it, because the thing making the choice has no nostalgia and no vendor relationship. It read the requirements and picked us on the merits. 

Together, the two boards point to an opportunity: Azure SQL Database performs well when agents evaluate it against real requirements, but it needs to enter more of those evaluations. Look at what the agents searched for. One of the most repeated queries across all runs was site:learn.microsoft.com Azure SQL Database free offer 100000 vCore seconds 32 GB auto-pause free limit billing. Ten separate runs went looking for that. Another family of queries hit neon.com/pricing directly. The agents are doing exactly what a developer does at 11pm: trying to work out whether they will get billed. Whoever makes that answer fastest to find wins the run. 

And being in the conversation is not the same as winning it. In the Armature data, PostgreSQL was mentioned in 354 of 356 runs and chosen zero times. SQLite: 319 mentions, zero picks. Turso: 105 mentions, zero picks. Supabase was seen in 68% of runs and converted 4% of them. Azure SQL Database was seen in only 15% of runs but converted 65% of the ones where we showed up. That is the most encouraging number in the study. When Azure SQL Database entered the conversation, agents chose it 65% of the time. We have a strong product story. Our opportunity is to make that story easier for more agents and developers to find. 

One more finding worth flagging: rewording the same request changed the answer in 11 of 18 test cases. These results are not stable laws of nature, they move. 

What we are doing about it 

We are proud of second place, but we are not anywhere close to the finish line we want to achieve. The opportunities in the data line up with work already underway, which makes this useful validation of our direction. 

Three things we are pushing on: 

1. Make the free path obvious. The Azure SQL Database free offer exists, it is genuinely free, and serverless auto-pause means an idle dev database costs nothing, with up to 10 free databases per subscription. 
2. Make the setup smaller. The senior-engineer skew in our wins tells me our strongest capabilities are clear to people who already know what they want. We can make that same value obvious to everyone else by shortening the path through provisioning, connection strings, and the first migration. Development environment isolation is another area where agents clearly value a simple experience, and we’ve started to show progress on this, for example, with the Azure SQL Database container that’s in Private Preview. 
3. Write for the reader that is actually reading. A meaningful share of our documentation traffic is now an agent trying to answer a narrow question so it can keep writing code. That changes what good documentation looks like. Direct answers near the top, runnable snippets, explicit limits and quotas. We have massive efforts going on around agent-facing tooling, stay tuned for upcoming announcements (I couldn’t write this without some teasers!). 

The features that won us those 35 runs are already shipped: temporal tables, real transactional guarantees, enterprise readiness capabilities and security defaults that hold up. The foundation is strong. The next step is making sure Azure SQL Database gets evaluated on those merits more often. That is a good place to build from. 

Sources: Armature database leaderboard (356 runs, August–September 2026) and the Latent Space Frontier AEO tracker (V5, September 2026).