Measuring agents honestly, a faster agent runtime, the CNCF radar, and a guard for critical DNS records
Two AI picks: LeadDev's three questions for measuring agentic AI, and AWS making its next generation AgentCore runtime the default (start times are vendor-reported). Two engineering picks: the CNCF Technology Radar for Q3 2026 on how platform ownership shapes AI infrastructure choices, and Route 53 record set protection, which blocks accidental changes to critical DNS records.
Microsoft says the fault was introduced during a network capacity upgrade. Treat capacity changes as risky deployments, with a staged rollout and a tested way back.
After the upgrade, a router overloaded a critical communication system. Schedule routine changes to shared dependencies for quiet hours and rehearse the rollback.
The petition site lists 5,965 current Nvidia employees among the signers. With leadership unsettled at a key GPU supplier, keep hardware and model plans flexible.
The regulator objected to a single app-wide authorisation covering a feature users were not clearly told about. Ask for consent per feature when it reads personal data.
Google claims it is 2.5 times faster than Flash 3.8 even in high mode; that is a vendor figure. Teams that built on Flash should rerun their own evals.
Ars Technica · 12 Nov 2026
AI & Agent Engineering
Engineering metricsverified
LeadDev: three questions to answer before you measure agentic AI
LeadDev asked developer experience researchers what to measure about agentic AI. The three questions: how developers feel, who owns what an agent builds, and how AI affects delivery. It cites surveys in which 41% of developers are hooked on AI coding, and 87% of organizations claim AI governance policies but fewer than a quarter have implemented controls. A McKinsey report put 60% of agentic AI spending on checking and correcting outputs.
Why it matters
The figures come from separate surveys and reports, so read them as signals, not benchmarks. The framing is the useful part: activity and volume show where the system strains, not the return. Laura Tacho of AWS suggests watching rework rate, code review, wait time and pull request size.
Do this week
Pick one rule about what agents may do without a person and check whether a pipeline gate enforces it or it is only written down. Then start a weekly baseline of rework rate, review wait time and pull request size.
AWS makes the next generation AgentCore runtime the default for new runtimes
New runtimes in Amazon Bedrock AgentCore now use the next generation runtime by default. It restores each session from a snapshot of the initialized agent and scales idle agents to zero. Each session still runs in its own Firecracker microVM. AWS reports session starts under 2 seconds at P75 on a 2 GB image, down from 29.7 seconds, and up to 68% less billed memory per session for multi-turn agents. Existing runtimes stay on the current version.
Why it matters
A default change means new runtimes behave differently from existing ones unless you set the version. The figures are AWS's own, for a 2 GB image. If the start time holds for your images, a separate microVM per session becomes practical for interactive agents, and scaling to zero changes what idle sessions cost.
Do this week
If you run agents on AgentCore, choose the version on purpose. Set platformVersion to V2 on one non-critical runtime, compare session start and billed memory with V1 for a week, then move the rest or keep V1 with a written reason.
CNCF published its Technology Radar for Q3 2026, produced with SlashData from a developer survey. It reports that 54% of developers with independently managed tooling use hybrid approaches for AI workflows, compared with 5% at organizations building custom internal platforms. 80% of developers surveyed rated Prometheus four or five stars for maturity, and 96% of current and former OpenTofu users are likely or highly likely to recommend it.
Why it matters
This is a survey of opinion, so it shows what practitioners trust, not what performs best. Two signals are still useful. Organizations that build a custom internal platform rarely mix approaches for AI workflows, so that choice gets made early. And OpenTofu's recommendation rate makes it a credible default for infrastructure as code.
Do this week
Write one paragraph on how your platform handles AI workflows today and who decided it. If nobody did, put it on the next platform review. Download the report and compare your observability and infrastructure as code tools with the radar.
Route 53 adds record set protection to block accidental changes to critical DNS records
Amazon Route 53 now supports record set protection. Administrators protect records by name and type directly in the hosted zone, and any change to a protected record is blocked until an authorized user removes the protection. One protection covers all records with that name and type, including weighted and latency records. Previously this needed record-level conditions in the IAM policies of every user and role that could change the zone.
Why it matters
A wrong change to an apex, MX or production record is an outage. Until now the guard lived in IAM policies spread across every user and role, and a blocked caller had no indication why. A protection in the zone can carry a comment of up to 256 characters and can be created before the record exists.
Do this week
List the records whose change would take production down: zone apex, MX and main production endpoints. Protect each with a comment naming the owner, and restrict the DeleteResourceRecordSetProtections permission to a small set of roles.
In one Route 53 hosted zone, pick the records whose accidental change would cause an outage, such as the zone apex, MX and production endpoints. Add a record set protection to each with a comment that names the owner. Then try a change to one of them from your normal deployment role.
Why. Route 53 blocks any change to a protected record until an authorized user removes the protection. Trying a change shows whether your pipeline reports the block clearly and who is able to remove a protection.
Measure. Record how many critical records are protected, how many roles hold the DeleteResourceRecordSetProtections permission, and whether the blocked change produced a clear error in the pipeline.
Classic / essential read
Cannot Measure Productivity
Martin Fowler · 5 min
Written in 2003, it explains why lines of code and function points fail as measures of output. The same argument applies to counting what agents generate: more code is not more output.
One lesson. Any true measure of software development productivity must be based on delivered business value, not on the volume of code or features produced.
Today's picks share one idea: decide on purpose instead of by default. Choose what you measure about agents, which runtime version you run, how your platform handles AI workflows, and which DNS records nobody should change by accident.
Get this every morning.
Free. Same issue for everyone. One click to unsubscribe.