Coding Agents Broke the Review Pipeline, Not the Code

AI-assisted teams produce 4x the code but defect rates spike from 9% to 54%. Factory pivots to software factories. Fable 5's free window closes today.

Three stories collided this week that, taken together, tell us something uncomfortable about the state of AI-assisted software engineering. The numbers are in, and they’re not what the demos promised.

The throughput trap: 180% more code, 12% more value

The data we’ve been waiting for arrived. Faros AI instrumented 22,000 developers across 4,000 teams and tracked what happened as adoption of AI coding tools ramped up. Devs merged more PRs. Throughput per engineer climbed. Then the other shoe dropped:

• Code churn up 861%. Code that gets rewritten or deleted within weeks of landing.
• Incidents-to-PR ratio up 242.7%. More code = more things breaking in production.
• Defect rates jumped from 9% to 54% in teams with heavy AI adoption, according to two large-scale studies covered by AI Insiders.

The punchline: teams produced roughly four times the code but captured only ~12% more delivered value. Volume swallowed quality.

Addy Osmani, engineering lead at Google, wrote the definitive take on June 15:

“The hard part of engineering moved from writing code to deciding whether to trust it, which makes review the most leveraged skill in software right now.”

He’s right. Code review used to work because a senior engineer could read code faster than a junior could write it. That ratio is now inverted. An agent produces a thousand lines of well-formatted code in seconds. Human reading speed hasn’t changed. The constraint moved downstream — to the one step nobody accelerated.

The same tools generating the extra code are the best tools for reviewing it. Osmani points Claude Code and Codex at incoming PRs to triage the queue. But that’s a workflow most teams haven’t adopted yet, and the Faros data shows the cost.

Factory goes full “software factory”

On June 15, Factory announced version 2.0, dropping the individual-developer framing entirely. The pitch: stop thinking about coding agents and start thinking about a self-improving system that ingests bug reports, customer feedback, and business requirements — then orchestrates the full SDLC from spec to deployment.

Factory co-founders Matan Grinberg and Eno Reyes frame it around three principles:

• Model independence. No single model fits every task. Swap between providers or let a router pick.
• Sovereign intelligence. Host it yourself. Feed every agent session, code review, and resolved incident back into the loop. The system gets smarter inside your walls.
• Continual learning. Every stage feeds every other.

It’s a smart pivot. The “coding agent” category is crowded with ~30 CLI tools, and the Faros data makes the individual-productivity pitch harder to sustain. Factory is betting the enterprise sale won’t be won on benchmark scores — it’ll be won on systems that close the loop between code generation and production outcomes.

Fable 5 free window closes today

Today, June 22, marks the last day of Anthropic’s 14-day free Fable 5 access for Claude subscribers. After today, Fable 5 moves to usage credits at 2x the rate of Opus 4.8.

The model itself remains suspended globally under the June 12 export control directive. Anthropic’s Chris Ciauri said restoration would come “within days” on June 18 — four days ago — but the API at claude-fable-5 still returns errors as of this morning. Kalshi traders price ~58–67% odds of restoration before July 1.

For teams that built agent pipelines on Fable 5: today is the deadline to either have a fallback model running or accept you’re paying credits for a model you can’t reliably access. If you’re still single-sourcing your agent’s LLM, the Fable 5 saga is your wake-up call. We’ve written about the multi-provider imperative in our agent infrastructure stack analysis — it’s no longer theoretical.

What we’re watching next week

• GPT-5.6 is in late-stage testing. OpenAI’s Chief Scientist called it a “meaningful leap” with a fix for reward hacking as the driver. Rumored 1.5M-token context window. Late June launch expected.
• Gemini 3.5 Pro has 9 days left on its self-imposed June deadline. Still in Vertex AI enterprise preview only.
• DeepSeek V4.1 rumored for June with visual multimodal capabilities. The 75% permanent price cut on V4-Pro announced May 23 keeps pressure on pricing.

The coding agent story is the one that matters most this week. The tools are real, the output gains are real, and the review pipeline is on fire. The teams that figure out agent-assisted review will ship faster and safer. Everyone else is just generating 861% more churn.