Agent sandboxes that do not idle, DR tests that find real gaps, and Node security releases due on November 11
Two AI picks: Google's Agent Substrate for idle-aware agent sandboxes, and Dropbox's plan, test and review process for delegating work to coding agents. Three engineering picks: AWS on four levels of disaster recovery testing, the Node.js security releases due November 11, and Cloudflare on post-quantum Privacy Pass.
AI & Agent Engineering
Agent infrastructureverified
Google's Agent Substrate hits 1.0 and snapshots idle agents so sandboxes do not hold a pod
Google says Agent Substrate, gVisor and OpenRL were accepted into the CNCF Sandbox, and Agent Substrate reached 1.0 and is GA on GKE. Agent state lives in Actors, execution in reusable Workers from a warm pool, and Gateways enforce access and inject secrets. Idle agents are snapshotted and restored in under 500 ms. Density, restore time and activation figures are vendor-reported.
Why it matters
Agents spend most of their time waiting on models, APIs or people, so a pod per agent wastes memory and ties identity to the pod. Separating agent state, identity and secret injection from compute is a design you can copy even without this project.
Do this week
Measure what share of your agent sandbox time is idle waiting. If it is high, write down how you would keep secrets out of the agent environment and what state you would snapshot, before choosing a runtime.
Dropbox: how coding agents were handed defined tasks with a plan, local tests and human review
Dropbox describes delegating defined tasks to coding agents. Each task started with a written plan agreed by engineer and agent, then the agent ran builds and tests locally and recorded results for review. A deferred test cleanup was split into 53 tasks and finished in one week, and 22 accessibility features were delivered. These are Dropbox's own reports.
Why it matters
The useful part is the process, not the speedup: plan first, rules stored with the code, recorded test results, and an engineer still accountable for the review. That is a template for delegating work without losing control of quality.
Do this week
Pick one long-deferred chore, such as a test cleanup. Split it into small tasks that each have a written plan and a pass or fail check, and give an agent two of them. Compare review time with doing them by hand.
AWS: four test levels find different disaster recovery gaps, and a planned failover alone misses most of them
Part 2 of an AWS series describes four validation levels: tabletop, planned failover, surprise drill and game day with fault injection. In one case study, Route 53 records still pointed to the primary Region 45 minutes into the test and recovery took more than twice the 5-hour target. Only one engineer could validate the databases, and finding the procedure cost 90 minutes.
Why it matters
A scheduled test proves the mechanism works, not that people can coordinate under pressure. Each level exposes a different class of gap, such as unowned steps, wrong DNS, or tribal knowledge.
Do this week
Check whether your recovery environment is pre-provisioned and whether account quotas in the recovery Region match production. Then book a tabletop this month and ask who declares failover and who backs them up.
Node.js pre-announces HIGH severity security releases for 26.x, 24.x and 22.x on November 11
The Node.js project says it will release new versions of the 26.x, 24.x and 22.x lines on or shortly after Wednesday, November 11, 2026. The highest severity issue fixed in each line is HIGH. The notice says end-of-life versions are always affected by a security release.
Why it matters
A pre-announcement gives you time to plan the upgrade. Teams on versions past end of life get no fix, so the only remedy is to move to a supported line.
Do this week
List every service and base image by Node major version today. Flag any not on 26, 24 or 22, and reserve time on November 11 to patch and redeploy the rest.
Cloudflare: Privacy Pass still uses RSA and elliptic curves, and post-quantum signatures are far larger
Cloudflare says Privacy Pass relies on classical cryptography, so a quantum attacker could forge tokens and bypass rate limits, though token unlinkability holds even then. It notes ML-DSA-44 signatures are 2,420 bytes against 64 bytes for ECDSA-P256, and explores a blind signature design called MOLE with tradeoffs in security assumptions.
Why it matters
The pattern is separate from TLS: any system that signs tokens, tickets or credentials has the same exposure and the same size cost. Forgery and privacy break differently, which affects how urgent each migration is.
Do this week
Inventory where you issue or verify signed tokens, such as API keys, JWTs and rate-limit passes. Note the signature algorithm and the size budget in headers or cookies, so you know what a roughly 38x larger signature would break.
Gather the owners of one critical service. Walk the recovery runbook step by step without touching production, and ask who declares failover, who owns each step and who is the backup.
Why. A tabletop is the cheapest DR test level, and AWS reports it surfaced three boundary gaps in one case study, including a step with no named owner.
Measure. Count steps with no named owner or backup, steps that depend on one person, and the estimated end-to-end time compared with your stated recovery target.
Classic / essential read
Microservice Trade-Offs
Martin Fowler · 15 min
It lists the benefits and costs of microservices side by side and insists the choice depends on your context. It reads well beside this week's stories on isolation and recovery.
One lesson. Treat architecture styles as trade-offs to evaluate for your team, not as improvements to adopt by default.
This week's common thread is verifying assumptions before they are tested for you: agent idle time, recovery quotas, runtime versions and signature sizes. Pick one this week and check it.
Get this every morning.
Free. Same issue for everyone. One click to unsubscribe.