Field note
Formula 1 and AWS described a system for onboarding new data sources that had previously taken six to eight weeks and contributed to an 18-month backlog across 12 sources.
A business requirements document lands in S3, triggers Lambda, and starts agents in Amazon Bedrock AgentCore. The workflow produces three pull requests covering ingestion and infrastructure, DBT transformations, and governance requirements including GDPR. A GitHub App and Jira preserve the normal engineering review path.
The control design is more important than the time claim. Skills are modular and bounded. Outputs pass through token and secret scrubbing, tool-output summarization, and a separate assessment pass before review. Schema evolution is event-driven rather than handled by one permanently open agent session.
AWS says code generation dropped to roughly 40 minutes, followed by hours for deployment and review, and that agents perform about 95% of the work autonomously. Those are case-study claims, but the architecture is reproducible: documents enter, typed artifacts and pull requests leave, and humans retain the merge boundary.