Patrick Debois: In AI Coding, Risk Is the Bottleneck, Not Review
bibryam · x · 2026-11-11
Patrick Debois argues that in an AI software factory — where agents generate changes, run tests, and open PRs faster than humans can review — the review queue only looks like the bottleneck. The real question is: under what conditions would we let a change ship without anyone reading every line?
He lays out four levers for managing risk:
- Reduce likelihood: tests, security/policy checks, static analysis, architectural constraints, adversarial/reviewer agents, isolated execution
- Verify outcomes: acceptance scenarios, coverage, before/after behavior, independent verification agents, evidence bundles linked to intent
- Locate the risk: rank by system criticality, customer exposure, blast radius, data sensitivity, reversibility
- Reduce impact: bulkheads, feature flags, circuit breakers, automatic rollback, rate limits, monitoring with auto-recovery
Changes should be routed by risk: low → ship, medium → canary, high → more verification, critical → human approval. His core thesis: AI factories won't become autonomous by perfecting review, but by making mistakes survivable.
More from coding & agent
- Natural language search over autonomous driving data with Cosmos-Embed1 + OpenSearch — 4310sy · 2026-11-11
- Spawning an 'interpretability researcher' subagent via a clever system prompt — Sauers_ · 2026-11-11
- Building a walkable 3D home from floor plans with Claude — full prompts shared — FinanceYF5 · 2026-11-11
- Two prompts run an entire AI-powered 3D real-estate tour business, from prospecting to delivery — FinanceYF5 · 2026-11-11
- Tektronix launches ~$500 12-bit DPO scope with AI MCP support in closed beta — IgorCarron · 2026-11-11
- Agent Bridge: Open-Source MCP Server Drives Your Logged-in Chrome, Cutting Token Use 57% vs Claude in Chrome — falconsw · 2026-11-11