Cursor migrates Brex services through reviewed PRs
Brex reports using self-hosted Cursor agents to migrate 180 microservices from Micronaut 3 to 4 in one month with one engineer. The earlier effort had moved about fifty services in a quarter with three engineers.
brex.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Migrate this Micronaut service using the approved skills. Build and validate it, repair failures at most three times, and prepare a draft PR with unresolved risks.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- Sofhia de Souza Gonçalves
- Original account
- Read the source ↗
- Published
- 22 Jul 2026
- Source link checked
- 15 Oct 2026
- Reported result
- Brex reports using self-hosted Cursor agents to migrate 180 microservices from Micronaut 3 to 4 in one month with one engineer. The earlier effort had moved about fifty services in a quarter with three engineers.
- Human involvement
- Brex engineers configured tools and eleven skills, reviewed draft PRs and retained all production merge decisions.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Trigger a generation agent from Linear to transform code, run builds and open a draft PR.
- Trigger a validation agent by PR label to monitor CI, deploy to development and run integration tests.
- Allow up to three repair attempts, then escalate failures; humans review and merge.
Brex reports using self-hosted Cursor agents to migrate 180 microservices from Micronaut 3 to 4 in one month with one engineer. The earlier effort had moved about fifty services in a quarter with three engineers.
The working pipeline separated generation from validation. A Linear ticket triggered an agent that combined deterministic transformations with contextual Kotlin fixes, ran builds and opened a draft pull request. A label then triggered a validation agent to monitor CI, deploy to development environments and run integration tests.
When failures appeared, the second agent could patch them up to three times. After that it stopped and posted a report for a human. Both agents drew on eleven focused skills instead of one oversized prompt. Dedicated cloud nodes and headless authentication kept the process running when laptops were closed.
Engineers retained review and production merge decisions. Some changes passed the existing tests but caused subtle runtime bugs under production traffic, exposing gaps in the validation coverage. The source’s thirty-two-fold per-person acceleration and roughly ninety-percent cost reduction are company comparisons, not controlled benchmarks. This is a reported migration deployment with bounded retries and documented escaped defects, rather than evidence that passing agent-run checks made every service safe.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: www.brex.com · Sofhia de Souza Gonçalves
Frequently asked questions
What did the agent do?
It transformed services, built code, opened draft PRs, validated CI and development deployments, and attempted bounded repairs.
What remained the human’s responsibility?
Brex engineers configured tools and eleven skills, reviewed draft PRs and retained all production merge decisions.
What happened after repeated validation failures?
The validation agent could attempt repairs three times, then stopped and posted a report for a human rather than retrying indefinitely.
More details
About this account
Cursor agents transformed service code, ran builds, opened draft PRs, monitored CI and development validation, and attempted at most three repairs before escalating.
Firsthand account opened and read. Outcome is source-reported, not independently reproduced. Suggested prompt written by Agents In Use. Some defects escaped existing tests and reached production; humans retained merge approval.
Credit: Sofhia de Souza Gonçalves
Source published 22 Jul 2026
Source checked 15 Oct 2026
Tools and topics
Linear · GitHub · Bazel · Kotlin · CI
Linear · GitHub · Bazel
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.