Fix CI before GitHub does, and ship ML features one at a time
Two AI picks: Meta's F3 Lightning, which separates release from execution for ML features, and a LeadDev view of the tech lead as owner of AI guardrails. Two engineering picks: GitHub's default block on pull_request_target from December 2, and Cloudflare's provably safe incremental DNS builds.
AI & Agent Engineering
ML systemsvendor-reported
Meta F3 Lightning: releasing ML features one at a time without losing shared compute
Meta's feature framework F3 serves more than 20,000 features per ads request by merging them into one graph, but that graph was also the release unit, so one invalid definition blocked everything. F3 Lightning compiles, validates, deploys and rolls back each feature independently. Meta says comparable features reach production about 3.6x faster with no added compute cost.
Why it matters
Early per-feature prototypes raised CPU use because features stopped sharing work. Meta separated the release boundary from the execution boundary to recover it. These results are Meta-reported.
Do this week
List what your deployments bundle only because they share runtime. If one bad change can block unrelated changes, test whether you can ship them separately and keep the shared execution path.
LeadDev: AI is shifting tech leads from coordinating work to setting guardrails
A LeadDev opinion piece by a tech lead argues that engineers with frontier models now plan across the whole stack alone, so up-front coordination matters less. The tech lead's work moves to design review, checking plans against work in flight, defining guardrails such as shared agent skills and tooling, and holding the architectural bar.
Why it matters
This is one practitioner's view, not a measured result. It still names a real change: low friction makes it easy to add one more quick change, so consequences get skipped unless review and guardrails catch them.
Do this week
For the next design doc written with AI help, require named reviewers from each affected domain and one check against in-flight projects. Note each case where an agent went off track and turn it into a skill or a CLI check.
GitHub will block pull_request_target by default in public repos on December 2
GitHub Actions will block pull_request_target events by default in all public repositories from December 2, 2026, unless an Actions event policy allows them. Enforcement was previously set for November 2. Two brownouts run on November 9 and November 16, 16:00 to 18:00 UTC. Private and internal repositories are not affected.
Why it matters
pull_request_target runs with the base repository's secrets, so running fork code there can leak them. Workflows that depend on it will fail during the brownouts and permanently after December 2.
Do this week
Before November 9, open Insights for each public repository to see which runs the rule would block. Allow only the workflows that truly need the trigger, and rework any that check out fork code.
Cloudflare cut DNS build memory 85% by building only what it can prove is safe
Cloudflare says its Zone Builder now gets about 700 DNS record changes per second, up from about 250 in 2022, and some zones hold over 20 million records. A planner picks one of five build types and falls back to a full build when unsure. Peak memory fell 85% and change-to-edge time halved.
Why it matters
The rule is that a targeted build must write keys identical to a full build, and uncertainty never reduces work. That makes an optimization safe to ship. These numbers are Cloudflare-reported.
Do this week
If you have an incremental rebuild path, add a test that compares its output byte for byte with the full rebuild, and default to the full path whenever the change event is unclear.
Pick one release unit in your system that bundles many items, such as a shared config or merged build. Write down how often one failing item delayed the rest in the last month, and sketch which items could ship alone.
Why. It shows whether bundling is costing you release speed, using your own incident history instead of a vendor figure.
Measure. Count of releases delayed by an unrelated failure in the last 30 days, and the median delay in hours for those releases.
Classic / essential read
Is Design Dead?
Martin Fowler · 40 min
It separates planned from evolutionary design and shows which practices make evolving a design safe. That question returns whenever AI makes code cheap to change.
One lesson. Cheap change does not remove design; it moves design into refactoring, tests and review habits that keep the structure clean.
Three checks fit this week: review your public repositories for pull_request_target before November 9, find release units where one failure blocks unrelated changes, and add a correctness comparison to any incremental rebuild path.
Get this every morning.
Free. Same issue for everyone. One click to unsubscribe.