The Horizon · Issue 2

AI doesn't fix a broken process. It runs it at 10x.

7 November 2026 · Hossam EL-Kashef

Welcome back to The Horizon. One big story, three strange ones, one tool worth trying, one thing to do on Monday.

The big one: speed moved the queue

Everyone is measuring how fast code gets written now, and fair enough. An agent writes in minutes what used to take a day.

But code written isn’t code shipped. Between the two sit review, tests, security checks, approvals and release. None of those got 10x faster. So the queue doesn’t disappear. It moves:

  • PRs pile up waiting for review.
  • Reviewers skim, because there are so many more of them.
  • Bugs slip through, and releases get bigger and scarier.

Google’s DORA team put it plainly in its 2025 year in review: AI improves throughput, “but often at the cost of stability if your foundation isn’t solid.”

Here’s the honest test I use. Take the AI away for a minute and look at the process. If reviews were already slow, tests flaky and ownership fuzzy, AI won’t fix any of that. It will give you a bigger mess, faster.

What I’d fix first:

  1. Small PRs. An agent will happily write 2,000 lines in one go. Don’t let it. Small changes get real reviews.
  2. Gates that scale. Tests, linters and security scans run in the pipeline on every change, not in someone’s calendar.
  3. One owner per change. Someone who can explain it and will answer for it. Agents propose, engineers decide.
  4. Measure flow, not output. Time from commit to production tells you more than lines written or PRs opened.

Speed is the easy part now. Flow is the job.

Three strange ones

  • AI agents booked four Uber rides. On its Q3 call, Uber said third-party AI agents booked just four rides and about 500 Uber Eats orders in a week. The CEO called it “very, very early”. The hype is ahead of the usage.
  • Apple put its engineers on an AI allowance. Bloomberg reports some engineers’ daily AI caps went from $300 to $200, with $100 under discussion. Token spend is now a budget line.
  • Glasses that type what you silently mouth. Subvocal opened preorders for glasses that read facial muscle movement and turn silently mouthed words into text, claiming 200+ words per minute. Not independently tested yet.

More in this week’s 8-story recap.

One tool worth trying

Toxiproxy, an open-source proxy from Shopify for simulating network trouble: added latency, timeouts, dropped connections. Put it between your service and a dependency in tests and you find out what your timeouts and retries actually do before production does it for you.

One thing to do on Monday

Search one service for outgoing HTTP or database calls with no timeout set. Some popular clients wait forever by default (Python’s requests, Go’s default HTTP client). Add a timeout based on that dependency’s real latency, and you’ve removed one silent way your system can hang.

See you next week.