Context
Seed-stage home management platform with a live beta and a growing product surface. The product runs on a multi-agent backend (RAG, knowledge graphs, document understanding), and the team had to ship faster without doubling headcount.
Problem
Six engineers covering a roadmap that would normally need twelve. The team's velocity was good. The runway for hiring was not. They needed throughput, not a quarter of recruiting.
What we tried first
Standard agile cadence with single-agent IDE help. The work moved, but the bottleneck didn't. Engineers weren't typing slower than they thought; they were taking on one problem at a time. The shape of the team was the constraint, not the speed of any one person.
What we shipped
Each engineer became the manager of an elastic AI team. Up to eight parallel Cursor agents per engineer, scoped by a senior, reviewed line by line. Project-specific Cursor rules. Planning templates. TDD enforcement that verifies tests can actually fail. Prior-art workflow that forces every agent to learn from what was built before. Frozen-golden eval suites for the agent backbone (RAG, knowledge graphs, document understanding). Sentry coverage. LangSmith traces on every agent path.
Outcome
89% faster shipping against the team's pre-agent baseline (internal tracking). 70% fewer defects in Sentry. 85% of production code AI-generated under engineer supervision. Acceptance rate on agent output around 79.6%. Peak week: 46 PRs merged with eight parallel agents. Inference cost about $9,500 per active week, published on purpose. Month 3 outpaced Month 1; the team gets better at this over time.
What we got wrong
Early on, agent context wasn't pinned tightly enough. One agent confidently rewrote a feature that was already in production. The fix was the prior-art workflow, plus a frozen-golden eval suite that runs on every PR. Once both were in, regressions surfaced before review, not after.