leadbybuild.ingArchive

LeadByBuilding Newsletter · 9 min read

Agents make you rethink identity, testing and write paths, and Node 27 may break your tests

Three AI picks: Cloudflare on how agent intermediaries hide users from sites, Shopify's headless architecture for testing mobile apps with agents, and Dropbox's sandbox, permission and undo design for a file-editing agent. Two engineering picks: GitHub's new Git storage design and the V8 changes in Node.js 27 Alpha.

AI & Agent Engineering

Agent identityverified

Cloudflare: agents are intermediaries, and sites lose sight of the user behind them

Cloudflare argues that remotely hosted personal agents act as intermediaries, so a site sees the agent and not the user. It lists problems this causes: shared accounts, friendly fraud, broken risk-based authentication, traffic commingling and resource exhaustion. It says CAPTCHA pass-through and broad access scopes do not fix them.

Why it matters

Blocking is no longer a policy once most traffic is agentic. If you run an API or a checkout, a ban on one agent also bans every user behind it. Cloudflare frames the need as identity, reputation and accountability per user.

Do this week

Review one sensitive flow this week, such as signup or checkout. List what you key rate limits and fraud rules on, and mark which of those signals disappear when an agent acts for many users.

Cloudflare Blog →
Agentic architectureverified

Shopify's Reflex lets agents test mobile apps without an emulator, with one test suite for iOS and Android

Shopify describes Reflex, an MVVM architecture where all app logic lives in a UI-free core module. Every screen exposes its state and actions as JSON, and a CLI drives them on macOS or the JVM. The code is written twice in Swift and Kotlin, and the same behavior tests run against both. Shopify reports 1,700+ behavior tests finishing in about four minutes on CI.

Why it matters

The design removes the slowest step in an agent loop, the simulator. It also replaces shared cross-platform code with a shared test as the parity mechanism. The test numbers are Shopify's own.

Do this week

Pick one screen in your app and ask whether its state and actions could be read as JSON without rendering. If not, list the imports that block it. That list is your cost estimate for headless agent testing.

Shopify Engineering →
Agent safetyverified

Dropbox built its file-editing agent around a sandbox, permission checks and version-history undo

Dropbox engineers describe merging two prototypes into one agent: a sandbox that runs installed software against files through rclone, and a desktop interface with permission controls. The agent's changes tie to version history for rollback. The UI uses the Codex App Server API, which sends JSON-RPC messages grouped into threads, turns and items. For broad requests the agent must propose a plan first.

Why it matters

The design lets an agent change real files because review and undo come from systems that already exist: permissions, revision history and previews. Gating file changes behind an accepted plan limits what a vague prompt can do.

Do this week

Before you give an agent write access to shared data, check that every change it makes can be traced to a conversation and reverted from existing history. If it cannot, make that a launch blocker.

Dropbox Tech →

Engineering

Distributed systemsvendor-reported

GitHub's new Git storage design orders only reference updates and keeps packs in Azure Blob Storage

GitHub details Project Rubicon. Git packs live in Azure Blob Storage, while references and the pack manifest live in a replicated log service called Ref-CAS that orders each reference independently. Virtual replicas are disposable compute with rebuildable disk caches, and compaction runs on separate nodes. In internal testing GitHub reports 5,000 concurrent agents sending 7,200 pushes per second to one repository.

Why it matters

The central point is that Git objects are immutable and content-addressed, so only the tiny reference update needs agreement. Splitting storage, serving and maintenance lets each scale alone. The throughput figures are GitHub-reported internal tests.

Do this week

When a write path is slow, ask which part truly needs ordering. Move everything else to immutable, content-addressed storage, and keep consensus to the small mutable pointer.

GitHub Blog →
Runtimesverified

Node.js 27 Alpha moves to V8 15.6, with changes to Math, Date and Array sort that can break tests

Node.js 27 Alpha upgrades V8 from 14.6 to 15.6. New features include Math.sumPrecise and iterator methods such as zip. Math functions now use LLVM libm, so some results differ in the last bit. Math.random changes its sequence for a given seed, getTimezoneOffset may return fractions, and large arrays sort with PowerSort instead of TimSort.

Why it matters

These are behavior changes, not just new syntax. Tests that hard-code floating point results, seeded random values or comparator call counts can fail on the upgrade. In the post's example, a sort comparator was called 19,220 times on v26 and 19,249 on v27.

Do this week

Run your suite against Node 27 Alpha in a non-blocking CI job now. Search it for hard-coded float outputs, seeded random expectations and comparator spies, and fix them before Node 27 reaches LTS.

Node.js Blog →

Worth experimenting with

Check which of your tests would break on Node 27 Alpha

Add a non-blocking CI job that runs your test suite on Node 27 Alpha. Collect failures and group them by cause: float results, seeded random values, timezone offsets and sort comparator behavior.

Why. Finds upgrade breakage months before LTS, while fixes are still cheap and you can report real regressions upstream.

Measure. Count of failing tests per cause, and the share that are test assumptions rather than product bugs.

Classic / essential read

CQRS (Command Query Responsibility Segregation)

Martin Fowler · 6 min

It explains splitting the model you update from the model you read, and it warns that for most systems this adds risky complexity. That pairs well with today's GitHub story about separating write ordering from reads.

One lesson. Use CQRS only where read and write needs truly diverge in a bounded part of the system. Fowler cautions against applying it across a whole system.

Read the original →

Commute watch · optional

Andrej Karpathy: From Vibe Coding to Agentic Engineering w/ Stephanie ZhanSequoia Capital · 29:49Become an Effective Software Engineering Manager • James Stanier & Gergely Orosz • GOTO 2023GOTO Conferences · 46:12

Agents change the assumptions behind your systems. Check what your identity and fraud rules depend on, make agent changes traceable and reversible, and keep consensus to the smallest piece of state that needs it.

Get this every morning.

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

Subscribe