leadbybuild.ingArchive

LeadByBuilding Newsletter · 8 min read

Sandboxing agent code, typed model decisions, and measuring developer friction instead of output

Three AI picks: Hyperlight micro-VMs for untrusted agent code, Spring AI's DecisionModel API, and Meta's LLDB design for debugging CPU and GPU code together. Three engineering picks: DeepL's Direct Connect drain with Cloud WAN, secret scanning for Actions logs, and LeadDev on measuring friction.

AI & Agent Engineering

Agent sandboxingvendor-reported

Hyperlight runs unmodified Linux programs in hypervisor-isolated micro-VMs, with snapshots for fast restore

A CNCF member post by a Microsoft engineer describes hyperlight-unikraft, which runs unmodified Linux programs in a micro-VM that a host application embeds. The guest has no network and no host files until the embedder grants them. In the post's benchmark, a numpy, pandas and scikit-learn script took 45 ms from snapshot against 1,119 ms native, while a Python hello world was slower at 22 ms against 17 ms.

Why it matters

Agents now run code nobody on the team wrote, so the blast radius of that code is a design decision. A hypervisor boundary is a stronger limit than a process sandbox. The timings come from one nested-virtualization VM and the project's own harness, so treat them as project-reported.

Do this week

List where your agents or plugins execute generated code today and what that code can reach. Pick one path and try a deny-by-default sandbox with explicit mounts and network rules, then compare start-up time with your current setup.

CNCF Blog →
Agent frameworksverified

Spring AI 2.1.0-RC2 adds DecisionModel for typed yes/no, choice and score answers with probabilities

Spring AI 2.1.0-RC2 adds a DecisionModel API. A prompt carries yes/no, choice or score questions, and the response returns the selected result with probabilities where the provider reports them. The release also streams tool call argument deltas for Anthropic, OpenAI, Amazon Bedrock Converse and DeepSeek, and adds a decision-based reranker, filter and tool index.

Why it matters

Many agent steps are small classifications that today parse free text. Returning probabilities lets the application set its own confidence threshold before it acts. The release notes say the answers are still not deterministic, and this is a release candidate, not a final version.

Do this week

Find one place where you parse a chat reply to branch logic, such as ticket triage or a guardrail. Prototype it as a typed decision with a high threshold and log the cases that fall below it.

Spring Blog →
AI infrastructureverified

Meta extends LLDB with an accelerator plugin so one session can debug both CPU and GPU code

Meta describes a common accelerator plugin architecture in lldb-server. The CPU and GPU are separate LLDB targets, each with its own architecture, address space and execution state, shown together in one LLDB or IDE session. A vendor plugin translates its device debugging API into processes, threads, registers and memory. GPU support is activated only when the app uses a supported GPU runtime.

Why it matters

AI workloads move between host code and device kernels, and bugs often cross that boundary. A vendor-neutral debugger interface reduces per-vendor tooling work. The post describes a design, so check which hardware your team uses is supported before planning around it.

Do this week

If your team debugs GPU code, note which debugger each hardware vendor requires today and where host and device sessions are split. Use that list to judge whether LLDB plugin support would remove a real workflow gap.

Meta Engineering →

Engineering & Leadership

Networkingverified

DeepL drains one Direct Connect link for maintenance with an AS_PATH prepend while other Regions keep local exits

DeepL connects Regions in Europe, North America and Asia to two European data centers through three Direct Connect locations and AWS Cloud WAN. The Direct Connect gateway picks paths by local preference, which is evaluated before AS_PATH, so the drain takes effect on each core network edge, which evaluates AS_PATH first. An on-premises prepend on one location moves only that Region's traffic.

Why it matters

Knowing which component makes the routing decision, and in what attribute order, decides whether a maintenance change does anything. The post also notes a cost: inter-Region data transfer while a link is drained.

Do this week

Before your next planned link or router maintenance, write down which component picks the path and which attribute it checks first. Test the drain in a lab or a low-traffic window and estimate the extra cross-Region transfer.

AWS Architecture Blog →
Supply chain securityverified

GitHub secret scanning now scans Actions workflow logs in public preview, with clear gaps

GitHub secret scanning now checks Actions workflow logs when a run completes. Findings appear in the repository alert list and link back to the log, and alerts remain after logs are deleted. It is in public preview for enterprises with GitHub Secret Protection or Advanced Security, and is disabled by default. Existing logs are not backfilled, and generic and custom patterns, AI-detected secrets and masked secrets are not detected.

Why it matters

CI logs are a common place for credentials to leak. The listed gaps and the lack of notifications during the preview mean this adds visibility but does not replace log hygiene or rotation.

Do this week

If you are on a qualifying plan, enable it on one repository this week and add a check of its alert list to your security routine, since there are no notifications yet. Rotate any secret it finds.

GitHub Changelog →
Engineering leadershipverified

LeadDev: measure developer friction alongside output, because faster coding moves the bottleneck downstream

A LeadDev opinion piece argues that output metrics such as pull requests miss time lost to slow pipelines, failed builds and cross-team waits. Its example is half an hour a day lost per developer, which it estimates at about 125 engineering hours a week for 50 developers. It also says that if AI cuts coding time by 40% while the pipeline stays slow, delivery will not speed up to match.

Why it matters

The numbers are the author's illustration, not a study. The useful idea is that AI-assisted coding raises the cost of slow builds, reviews and unclear failures, which output dashboards do not show.

Do this week

Pick your team's slowest recurring wait, such as CI queue time or flaky reruns, and measure it for two weeks. Put it next to your output metrics in the next planning review.

LeadDev →

Worth experimenting with

Measure one week of pipeline friction on one team

Choose one team and one recurring delay, such as CI queue time, flaky test reruns or waiting on a platform team. Record minutes lost per developer per day for two weeks, from CI data or a short daily tally.

Why. It gives you a number for friction that output metrics miss, which helps you argue for fixing the slowest step first.

Measure. Total minutes lost per developer per week and the share caused by the single worst step. If the worst step is under 5 percent of the total, pick a different delay.

Classic / essential read

Semantic Diffusion

Martin Fowler · 3 min

Fowler describes how a coined term loses its meaning as it passes from group to group, and says this tends to happen during the hype phase of an idea. Terms like agent and sandbox are in that phase now.

One lesson. When a term drives a purchase or design decision, go back to the people who defined it and check the original definition before you adopt it.

Read the original →

Commute watch · optional

DEF CON 34 - Main - The Sandbox is a Suggestion Deconstructing AI Agent Sandboxes - Elad MegedDEFCONConference · 42:43Reliability Engineering Mindset • Alex Ewerlöf & Charity MajorsGOTO Conferences · 27:11

Ask what each piece can reach and what it costs the people around it. Limit what agent code can touch, know which component decides a route, and measure friction, not only output.

Get this every morning.

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

Subscribe