How to Build an AI Software Factory: Source, Sink, Drain Across 921 Tickets
TejasKumar_ · x · 2026-11-07
IBM AI engineer Tejas Kumar published a deep dive on how he actually runs his "AI software factory" — agents working through a queue of tickets — distilled into a three-part mental model: source, sink, drain.
- Source: the human decides what gets built, written into GENESIS.md;
- Sink: an agent called pm turns it into issues on a Linear board — the queue (Daiku's board grew from 475 tickets on Oct 16 to 921);
- Drain: pm spins up as many coding agents as the token quota allows, merging to main and shipping.
He explains how urgent tickets jump the queue (pm spawns a dedicated agent, then the drain returns to capacity), why the level-4 agentic coding bottleneck "I lose track of my own project" was exactly the missing sentence, and what role remains for humans once agents do the rest. He also traces the term's history and why vendor definitions (Cortex, LaunchDarkly) never tell you the shape of the thing.
More from coding & agent
- Three ways to build an AI app: single agent, agent team, or graph workflow — rseroter · 2026-11-07
- New agent harnesses burn serious tokens to find their first bug, limiting who even tries — kuza55 · 2026-11-07
- Curated Bend 2 package ecosystem tracker snapshots 1,036 packages on BendHub — SubtleGradient · 2026-11-07
- AI auto-exploitation hits 91.3% on real codebases — vibe coding's security wake-up call — lxfater · 2026-11-07
- Matt Pocock: Opus 5.5 one-shots a 400px menu bug from a zero-detail prompt — mattpocockuk · 2026-11-07
- Qwen 3.8 freehand-draws a dog in MS Paint via computer use, and the internet loves it — Porespellar · 2026-11-07