Skip to content

AI Inflection Point (November 2025)

Category: concept Last updated: 2026-04-03 Status: draft

Summary

In November 2025, GPT 5.1 and Claude Opus 4.5 crossed a reliability threshold that entities/simon-willison calls the "inflection point": coding agents went from producing output that "mostly works" (requiring close supervision) to output that "almost always does what you told it to do." This shift unlocked truly autonomous agentic workflows and triggered widespread realization among software engineers returning from the holidays that AI coding had crossed a qualitative threshold.

Details

What Changed

Prior to November 2025, coding agents would produce code that was often buggy or only partially correct — usable, but requiring constant attention. The threshold crossing was not a sudden step change in benchmark scores but a subjective reliability shift:

  • Before: Agents produced code that mostly worked; you had to watch them closely
  • After: Agents almost always do what you told them to do — "which makes all of the difference in the world"

This was the result of Anthropic and OpenAI spending 2025 focused almost entirely on code, using reinforcement learning and the "reasoning models" pattern (first introduced by OpenAI's o1 in late 2024).

Why It Happened

  • 2025: Anthropic and OpenAI redirected most training resources toward code — reasoning models are particularly effective for code because correctness is verifiable
  • Reinforcement learning on code (where "right or wrong" is clear) drove rapid improvement
  • Claude Code (launched Feb 2025) validated that people would pay $200/month for coding tools, incentivizing continued investment

Impact on Engineers

Many software engineers took time off over the 2025 holidays and returned to find the technology had qualitatively changed. The realization spread in January–February 2026:

  • Engineers began producing 10,000 lines of code per day that "mostly work"
  • ==The bottleneck shifted from writing code to everything else: spec-writing, code review, QA, design decisions, and understanding what to build==
  • Prototyping became essentially free

The Dark Factory Pattern

One consequence of the inflection point is the viability of the dark factory pattern: software built without any engineer reading the code. Named by analogy to automated factories that can run with the lights off because no humans are present.

StrongDM (a security access management company) pioneered this starting in August 2025:

  • Policy: Nobody writes code; nobody reads the code
  • QA: A swarm of AI agent testers simulating real employees 24/7 in a fake Slack channel, making requests and triggering the software; ~$10,000/day in token costs
  • Infrastructure: Built their own simulated versions of Slack, Jira, Okta etc. from public API documentation using coding agents, avoiding rate limits
  • Insight: If you can't read the code, the QA challenge becomes "how do you know if software is good?" — answered through comprehensive simulation rather than code review

The OpenAI Codex team ran a parallel but broader experiment (published March 2026), building an internal product with 0 lines of manually-written code starting August 2025: ~1M lines of code, ~1,500 PRs merged, 3–7 engineers, avg 3.5 PRs per engineer per day, built in ~1/10th the estimated hand-coded time. See concepts/harness-engineering for full details.

The dark factory is named by analogy to automated manufacturing facilities that can run in complete darkness because no humans need to be present on the floor.

What Remains Human

Even with the inflection point crossed, Simon Willison argues experienced engineers are more valuable, not less:

  • Amplifying 25 years of experience with agents produces qualitatively better output than a novice using the same tools
  • The "prompt in one sentence vs. this is a hard problem" judgment — knowing which tasks are trivial for agents — requires domain expertise
  • Knowing when code is "done" (not just passing tests) still requires judgment
  • Cognitive load from managing 4 parallel agents is exhausting; experienced engineers manage this better

==Who is most at risk: Mid-career engineers — not beginners (who onboard faster) and not senior engineers (whose expertise is amplified), but those in between who haven't accumulated enough experience to direct agents effectively and aren't fresh enough to adapt natively.==

Predictions

  • entities/simon-willison: 50% of engineers writing 95% AI-generated code by end of 2026
  • Lenny Rachitsky's job market data (early 2026): record-high number of open engineering and PM roles at tech companies despite layoffs, suggesting demand is growing not shrinking

Key Claims & Data Points

  • November 2025: GPT 5.1 + Claude Opus 4.5 crossed the coding agent reliability threshold — [source: wc8FBhQtdsA]
  • StrongDM ran dark-factory coding from August 2025 with ~$10,000/day in QA token costs — [source: wc8FBhQtdsA]
  • 95% of Simon's code is AI-written as of early 2026 — [source: wc8FBhQtdsA]
  • Mid-career engineers identified as most at risk; new engineers onboard faster with AI — [source: wc8FBhQtdsA] (citing Thoughtworks engineering VP survey)
  • Cloudflare and Shopify hired ~1,000 interns in 2025; AI cut onboarding time from a month to a week — [source: wc8FBhQtdsA]
  • Prediction: 50% of engineers writing 95% AI code by end of 2026 — [source: wc8FBhQtdsA]

Open Questions

  • What does the quality gap look like between "most of the time works" and "all of the time works" — what's still missing? (raised by: concepts/ai-inflection-point, 2026-04-03)
  • How is the dark factory pattern adopted outside of security-adjacent companies where testing is easier to simulate? (raised by: concepts/ai-inflection-point, 2026-04-03)
  • Will the prediction of 50% of engineers writing 95% AI code by end of 2026 materialize — and how do we measure it? (raised by: concepts/ai-inflection-point, 2026-04-03)
  • How does the inflection point expand to non-code knowledge work (law, medicine, journalism)? (raised by: concepts/ai-inflection-point, 2026-04-03)

Sources