leadbybuild.ingArchive

LeadByBuilding Newsletter · 8 min read

Give agents tools, not buttons, and measure what AI rework costs

Three AI picks: WebMCP tools for storefront agents, a survey putting AI rework at 139 hours per engineer a year, and a lab for sharing GPUs across tenant clusters. Three engineering picks: Cloudflare's chunked range caching, post-quantum key wrapping for regional keys, and the TypeScript 7.1 beta.

AI & Agent Engineering

Agent interfacesvendor-reported

Cloudflare and Commerce.com turn on WebMCP for almost 100,000 storefronts, with a per-merchant opt-out at the edge

Cloudflare says WebMCP, an experimental browser standard from Google, gives agents structured page tools such as search_products instead of clicking through HTML. Its edge implementation injects a script automatically. For Commerce.com it used a Configuration Rule that reads custom hostname metadata, so any merchant can opt out without extra latency.

Why it matters

Agents that drive sites through HTML break on theme changes and cost many tokens. Exposing named tools is sturdier, but WebMCP is still behind Chrome flags and origin trials, so this is early. The opt-out rule is a reusable pattern for rolling out a platform-wide feature.

Do this week

If agents shop or operate in your product, list the five actions they should call and sketch them as tools. Prototype one behind a flag, and add a per-tenant kill switch before any wide rollout.

Cloudflare Blog →
Engineering managementvendor-reported

A survey puts AI rework at 139 hours per engineer a year, and half say new dependencies can ship unreviewed

LeadDev reports an ActiveState survey of over 300 security, platform and engineering leaders at organizations with 1,000 or more employees. 43% say at least 10% of an engineer's time goes on fixing or revisiting AI-assisted code, which the report turns into a floor of 139 hours per engineer a year. Half say new dependencies can reach production with no prior review.

Why it matters

The survey is vendor-run and self-reported, so treat the 139 hours as a prompt to measure, not a benchmark. The useful point is that rework hides in sprint carryover and unplanned tickets, while dependency growth driven by AI assistants outruns after-the-fact scanning.

Do this week

Add a rework label for AI-assisted work in your tracker for one sprint and count it. Separately, check whether adding a dependency has an approved path that is faster than going around it.

LeadDev →
GPU platformverified

A lab shows per-team Kubernetes control planes with fractional GPU sharing, built on simulated hardware

A CNCF post combines Kamaji, which runs each tenant's control plane as pods in one management cluster, with HAMi, which lets pods request GPU memory in MB and compute in percent. Using HAMi's mock device plugin, the author ran two isolated tenant clusters, and three pods each asking for 20,000 MB and 30% shared one simulated A100.

Why it matters

It tests the risky seam: the tenant API server must call HAMi's webhook back through the Konnectivity tunnel, and that worked. The author is clear it proves wiring and scheduling only, with no real GPU limits, performance or noisy-neighbour data.

Do this week

If teams ask for dedicated GPU clusters, reproduce this lab on a laptop first. Then repeat it on one real GPU node and measure memory limit enforcement and neighbour interference before promising shared GPUs to anyone.

CNCF →

Engineering

Caching and CDNvendor-reported

Cloudflare replaces tuned range sizes with fixed cache chunks and says R2 p99 time to first byte roughly halved

Cloudflare's old range caching mapped requests to a per-zone range size, and R2's 64 MiB setting could cost 128 MiB of transfer for a small read. Its Pingora cache already stores files in fixed-size chunks, so ranges now align to chunks. Cloudflare reports R2 p99 time to first byte dropped about half and tiered cache traffic fell about 6%. It ships as Origin Range Requests on all plans via a Cache Rule.

Why it matters

The design lesson is to use an existing storage unit instead of exposing a tuning knob. It also validates the entity tag and total size across subrequests, and serves the full response when a range cannot be honoured safely. The numbers are Cloudflare's own.

Do this week

If you serve large files or package artifacts through a CDN, look for a configured range or slice size. Test a seek-heavy workload with the new setting and compare p99 time to first byte and origin egress.

Cloudflare Blog →
Security architecturevendor-reported

Cloudflare adds hybrid ML-KEM-768 and X25519 protection to GeoKey Manager, with regional root keys that fail closed

Cloudflare says GeoKey Manager now uses hybrid ML-KEM-768 and X25519 with two key types: a symmetric regional root key per region and application-specific key pairs derived from it. Wrapped private keys are distributed globally, while root keys go through a separate provisioning service. Edge services refuse to start, and requests fail, if a required regional key is missing.

Why it matters

The design separates globally distributed ciphertext from regionally held keys, so data residency rests on key availability rather than on routing. It also protects private keys against harvest now, decrypt later without waiting for post-quantum attribute-based encryption.

Do this week

If you hold residency or key custody commitments, list where keys can be unwrapped. Check whether a missing regional key makes your service fail closed, and add post-quantum key wrapping to your crypto inventory.

Cloudflare Blog →
Language toolingverified

TypeScript 7.1 beta adds import attributes to ambient module declarations and es2026 as a stable target

Microsoft announced the TypeScript 7.1 beta, installable with npm install -D typescript@beta. Ambient module declarations can now carry import attributes, so a declaration for text or CSS can require with type text or type css. The beta also supports es2026 as a stable target and lib option.

Why it matters

Wildcard declarations such as *.md could not model attributes, so a forgotten with clause compiled fine and failed at runtime. Typed attributes catch that at build time, which helps if you import CSS, text or bytes through bundlers.

Do this week

Install the beta in a branch of a repo that uses wildcard module declarations, run the type check, and note which declarations can adopt attributes before the stable release.

TypeScript Blog (Microsoft DevBlogs) →

Classic / essential read

Circuit Breaker

Martin Fowler · 5 min

Remote calls can hang, and many callers waiting on one unresponsive supplier can exhaust resources and cascade across systems. Fowler describes the breaker that stops calling after a failure threshold.

One lesson. Wrap risky remote calls with a failure threshold and an alert when the breaker trips, so failure stays local instead of spreading.

Read the original →

Pick one thing to measure this week: how long an agent takes to reach a working build, or how many hours go to reworking AI-assisted code. A number beats a debate about speed.

Get this every morning.

Free. Same issue for everyone. One click to unsubscribe.

Subscribe