Working Surfaceissue · 12 Nov
Issue · 3 links
Issue

Thursday
12 November 2026

The day's three accounts come from engineers whose AI agents write most of the code, at Amazon Web Services, at Thoughtworks and on a Snowflake finance project. Each describes engineers putting their effort into tests, release checks and protected files that decide whether the agents' work is correct.

Swipe for 02 to 03
12 Nov
The note

Three engineering accounts published on 11 November argue that when AI agents write the code, the engineer's main work becomes writing the checks that decide whether it is correct. Each one keeps those checks out of the agents' reach, in tests, data files or a release pipeline that people control.

Amazon Web Services (AWS), Amazon's cloud computing business, released AWS Emulator, a free local copy of more than 200 of its services for testing software. Swami Sivasubramanian (12 Nov · 01), one of the principal architects of AWS's data strategy, explained how it was built in a guest post on the blog of Amazon's chief technology officer. AI agents wrote all of its code, working from end-to-end tests that they first corrected until the tests passed against the real AWS cloud. Every change must also pass a merge gate, a set of automated checks that includes several security agents looking for ways to attack it. Sivasubramanian concludes that engineers now decide what gets built and when the work is correct; the piece is the company's account of its own launch.

Matteo Vaccari (12 Nov · 02), a developer and Technical Principal at the consultancy Thoughtworks Italia, interviewed 26 developers at an organisation that relied heavily on AI coding agents. They told him they no longer trusted their tests, because the agents rewrote the tests along with the code. Writing on martinfowler.com, Vaccari proposes that developers act as the customers of AI-written code. They write customer tests that state inputs and expected results in plain terms, keep them as data in text files, and protect those files from changes by the agents. An AI agent then writes the small program that runs the files.

Sanket Prabhu (12 Nov · 03), an engineer with about ten years of building data platforms, built an AI agent that answers finance questions on Snowflake, a cloud data platform. Building the agent took a week, but moving it safely from a test environment into production took longer. He put its instructions, tool descriptions, finance data definitions and access rights in one code repository. They are released through the same approval gates as every other change. After every release, a fixed set of finance questions checks both the answers and which tool the agent chose. Most bad answers came from the data definitions, so people who knew those definitions reviewed them in the same pull request.

What they add up to

In all three accounts, the agent writes the code quickly and the slow, human part is defining what correct means. The check that matters sits somewhere the agent cannot quietly change: the real AWS cloud, a protected test file or an approval gate. That such checks need a named person who owns them is an inference, not any author's claim.

The case against

Each method depends on a reference that is easy to name: a real service to copy, a clear customer rule or a known finance question. Much product work has no such reference, and Vaccari himself says prototypes can skip this machinery while a team is still learning.

Receipts
All Things Distributed
AWS Emulator and the engineer's job when agents write the codeCompany story · 12 Nov 2026
Swami Sivasubramanian · 11 November 2026

AWS built its new local emulator of more than 200 cloud services entirely with AI agents, and its engineers' work moved to the tests and merge checks that judge it.

martinfowler.com
Tests as data: Ensuring AI-generated code does what we wantHow-to · 12 Nov 2026
Matteo Vaccari · 11 November 2026

Matteo Vaccari of Thoughtworks proposes that developers write plain-language customer tests, kept as protected data files, as the contract that AI coding agents must meet.

Snowflake Builders Blog
Agents Are Code Too: What I Learned Shipping a Cortex Agent Through Azure DevOpsHow-to · 12 Nov 2026
Sanket Prabhu · 11 November 2026

Engineer Sanket Prabhu released a finance AI agent on Snowflake through the same versioned, approved pipeline as other production changes, and tested it after every release.