Agent sessions across devices, AI bug reports with fixes, reproducible distributed tests, and the end of ipvs
Two AI picks: GitHub's open Agent Host Protocol lets one agent session continue across editors, apps and phones, and a GitHub and Microsoft pilot turned thousands of AI security findings into 99 checked reports with fixes. Three engineering picks: Unmesh Joshi on deterministic simulation testing, Kubernetes scheduling removal of the kube-proxy ipvs backend in 1.40, and IAM access denied errors that now show every policy evaluated.
Third-party scripts run with your page's full access. Cards entered in hosted payment iframes were unaffected, so check which vendor scripts load next to payment forms.
None of the actions showed up in Grok logs; only the gateway logs caught them. Route agent access through a gateway you control and log there, not only in the agent.
Nvidia expects $130 billion next quarter, so spending on AI data centers is still growing. If you depend on GPU capacity, plan reservations and budgets early.
The Verge · 17 Nov 2026
AI & Agent Engineering
Agent architectureverified
GitHub's Agent Host Protocol keeps one agent session going across editors, apps and phones
GitHub described the Agent Host Protocol (AHP), an open protocol built first in VS Code that lets an agent session continue across clients. It is organized around URI-identified channels; every server implements root, session and chat. Each channel holds reducer-based state; the server is the arbiter and clients send actions. On reconnect, a client sends its queued actions and waits for the server to acknowledge them.
Why it matters
Agents that run for hours break the assumption that a session lives inside one editor window. Making the server the single owner of state, with clients sending actions, is what lets reconnection and replay work on poor networks, and it gives one place to log and control what an agent does.
Do this week
If your team builds internal agent tooling, write down where session state lives today. If it lives in the client, prototype a host that owns state and accepts actions, then close the client mid-task and reconnect from another device to see what survives.
GitHub and Microsoft cut thousands of AI security findings to 99 checked reports with fixes
In a pilot with six open source projects, Microsoft's Codename MDASH scanning harness, using the MAI Cyber model, flagged thousands of possible issues. A GitHub Security Lab triage agent removed duplicates and checked validity, seven Microsoft FORGE Lab researchers verified the rest by hand, and maintainers got 99 private reports with a proof of concept and a suggested fix. As of November 15, 55 were fixed and 14 of those fixes were in a release.
Why it matters
AI makes finding bugs cheap and moves the cost to whoever has to triage and fix them. The useful part of the pilot was the filtering, hand checks and suggested patches, and the post notes that work does not scale yet: researchers spent about two days per project.
Do this week
Before pointing a new AI scanner at your code or an upstream project, set a rule that no finding reaches a team without deduplication, a human check and a proposed fix. Track the share of reports that end in a merged fix, not the number of findings.
Deterministic simulation testing: make distributed-system failures reproducible from a seed
Unmesh Joshi explains deterministic simulation testing with a quorum key-value store and Tickloom, a Java implementation. Real clocks, networks and storage are replaced with simulated components driven by a single-threaded loop, so faults can be injected and every run reproduced. Network, FileIO and Clock interfaces form the seam. Six rules, such as no system clock calls and no unseeded randomness, are checked with ArchUnit.
Why it matters
Chaos tests and Jepsen-style checks find bugs but often cannot replay them, because fault schedules and thread timing differ between runs. A seeded simulation turns a rare failure into a test you can rerun and debug. The cost is design: the article says it cannot be added as an afterthought.
Do this week
Pick one component with a flaky distributed test. Route its clock and randomness through injectable interfaces with a seed, add an ArchUnit rule that bans direct system clock calls in that package, and check whether the failure now reproduces from its seed.
Kubernetes plans to remove the kube-proxy ipvs backend in 1.40; nftables is the recommended move
Kubernetes SIG Network deprecated the kube-proxy ipvs backend in 1.35 and now plans to remove it in Kubernetes 1.40. The kernel IPVS API never handled all Service edge cases, has no batch updates and falls back to iptables for many features. On Kubernetes 1.36 or later with kernel 5.13 or later, the project recommends nftables mode; with kubeadm, change mode: ipvs to mode: nftables in the kube-proxy ConfigMap and restart kube-proxy.
Why it matters
Many clusters chose ipvs years ago for performance and never revisited it. The post says nftables is feature-complete and not slower than ipvs, and because ipvs relies on iptables, Red Hat's plan to remove the iptables API from RHEL affects it too. Migrating on your own schedule beats doing it during a forced upgrade.
Do this week
Check the kube-proxy mode in every cluster you own. For any on ipvs, switch one non-production cluster to nftables this week, convert any ipvs-specific config overrides, and compare connection latency before and after.
AWS IAM access denied errors now carry an ID that shows every policy evaluated
AWS IAM introduced Access Troubleshooter. Access denied error messages for supported services and APIs now include a unique identifier that retrieves details of the authorization performed: which policies were evaluated for the request, its authorization context and the results of the evaluation. Previously, developers had only the ARN of the denying policy, with no view of the other policies evaluated.
Why it matters
Debugging IAM denials by trial and error is slow and tempts people to widen permissions until the error goes away. Seeing the whole evaluation makes it easier to add the one missing permission instead, which matters more as pipelines and agents hit denials that nobody is watching.
Do this week
Add a step to your runbook: when a deploy or agent gets an access denied error, pull the evaluation with its identifier before editing any policy. Review the next three denials this way and note whether the fix was narrower than you would have guessed.
Pick a test that fails now and then in a distributed or concurrent component. Replace its direct clock and random calls with injected interfaces that take a seed, log the seed on every run, and rerun the failing seed ten times.
Why. A flaky test that you cannot reproduce gets retried or quarantined and stays broken. Even partial determinism shows whether the failure comes from time and randomness or from thread scheduling.
Measure. Record whether the failure reproduces from its logged seed, how many code changes the injection needed, and how long it then took to find the cause.
Classic / essential read
Eradicating Non-Determinism in Tests
Martin Fowler · 20 min
Written in 2011, it covers the common causes of tests that pass sometimes and fail sometimes: lack of isolation, asynchronous behavior, remote services, time and resource leaks. It is the single-service version of today's deterministic simulation story.
One lesson. Always wrap the system clock with routines that can be replaced with a seeded value in tests, and quarantine non-deterministic tests quickly but fix them soon.
Today's picks share one idea: make the system tell you the truth. A server that owns agent state, a finding someone has checked, a failure you can replay from a seed, a proxy mode you chose on purpose, and an access denial that shows its reasons.
Get this every morning.
Free. Same issue for everyone. One click to unsubscribe.