
RPI, Specs Drift, and Compaction

Context engineering with Dex Horthy
Topic 7 of 10
Published Jul 15, 2026 · Clip 1:01:03–1:13:50 (12:46)
Dex explains the original research-plan-implement (RPI) workflow from August 2025: parallel sub-agents read code and compress it into a markdown doc (100k tokens → 10k doc), then a new session plans, then implements. Planning became popular as a lever to get agents to work longer, anchored by a plan document. But the original plans were anti-leverage: reviewing both plan and PR doubled reading time, and specs drift from code — the perennial spec-driven development problem (he cites a year-old Spec Kit GitHub issue). The docs became disposable: tokens are cheap, time is expensive, so research is recreated live each time. He then covers intentional compaction as a context engineering building block, warns against horizontal plans (database → services → API → frontend) versus vertical slices a human would build, and defines the smart zone (~first 100k tokens, 200k for big models) versus the dumb zone, plus the 'you're absolutely right' signal and the four context dimensions: size, quality, missing information, and trajectory.