Anecdotally: Our AI productivity gains for production-grade code have increased by orders of magnitude once we got our autonomous XP/TDD agents working in @claudeai Claude Code.
Our sub-agents represent existing roles and responsibilities that we already had established and practiced. So writing them down required almost no effort.
Getting the main agent loop (CLAUDE(.)md) to orchestrate the agents took about a week to get right.
In total we have:
code-quality-reviewer
dependency-manager
locale-translation-specialist
test-coverage-analyzer
code-refactorer
frontend-design-specialist
qa-tester
db-migration-specialist
green-minimal-implementer
red-test-writer
Attached is a snapshot of how that looks in practice.
New Anthropic research: Estimating AI productivity gains from Claude conversations.
The Anthropic Economic Index tells us where Claude is used, and for which tasks. But it doesn’t tell us how useful Claude is. How much time does it save?
