Running an agent fleet across 15 repos: the bottleneck turned out to be the human
yi_ding · x · 2026-11-19
During his sabbatical, Rahul Parundekar ran a fleet of agents across 15 repos—a design system, auth service, template repo, customer apps, and consulting projects—and measured everything. His biggest surprise: the slowest part of the whole software factory was himself.
He published the findings as an ebook, Architecting a Software Factory, Part I. A software factory isn't a product but a way of building software (like waterfall or agile), organized into five flows:
- Direction: what the factory builds—customer requests, production alerts, even the agents' own requests
- Build: agents messaging each other, pulling in outside context, planning and writing code
- Delivery: agents reviewing code against instructions, shipping it
A measured, first-hand multi-agent engineering retrospective with plenty of unexpected learnings.
More from coding & agent
- Same game prompt: Astra ships for $17 vs Fable's $65 — 3x cheaper, fewer bugs — heypearlai · 2026-11-19
- GPT-6.2 Astra Costs 3x Less Than Fable 5.2 but Wins on Vegetation and Fire — heypearlai · 2026-11-19
- Same 5.5h Compute, Wildly Different Output: Astra 32.9MB vs Fable 242KB — heypearlai · 2026-11-19
- Same Prompt, Same Game: Astra Built It for $10.03 vs Fable 5.2's $22.78 — heypearlai · 2026-11-19
- Mastra Launches Factory 'Software Factory,' Auto-Merged 238 PRs Last Week — massimosgrelli · 2026-11-19
- Firecrawl Local Goes Open Source: 37x Faster Rust Scraping, 11MB Idle Memory — gaganghotra_ · 2026-11-19