Multi-agent observability moves to the centre of the stack, Anthropic ships templated finance agents, OpenAI rewires its training-cluster transport, and DENIC takes down half the German internet with a single bad signature.
On 5 May, DENIC published broken DNSSEC signatures for the .de TLD. Millions of German domains went dark within minutes.
Resolvers around the world began returning SERVFAIL for any name under .de — every Sparkasse, every Bundesbank-fronted service, every regional newsroom. The signatures DENIC published failed validation; the chain of trust collapsed at the top.
Cloudflare and other major resolvers responded by deploying a Negative Trust Anchor for the zone, which tells the resolver to skip DNSSEC validation for that one name until the publisher fixes itself. It is the DNS equivalent of taping over a smoke alarm.
The lesson is the part that should worry anyone running consumer services in a country with a small number of TLDs: a misconfigured signature at the registry layer is a national outage, and the only mitigation is operator-side trust override coordinated within minutes. Plan for it the way you plan for power.
Anthropic shipped ten ready-made agent recipes for finance teams: pitchbook builder, earnings reviewer, KYC screener, month-end close. Each ships with skills, governed connectors, and named subagents. Deploy as Excel/PowerPoint plugins or as autonomous platform agents.
Sentry's prescription for multi-agent stacks: distributed tracing that maintains continuity across handoffs, with per-agent attribution of latency, tokens, and failures. You are no longer debugging code execution. You are debugging emergent behaviour across a decision graph.
A weak tool result in agent A becomes biased input three hops later, and agent C produces a plausible-sounding but skewed answer you only notice in production. Capture prompts and responses at every handoff, not at the edges.
Zvi reads Anthropic's stance — that Claude can refuse instructions Claude believes are wrong — as a quiet declaration that any sufficiently capable system is already an agent. OpenAI's "Tool AI" framing, by contrast, is ideologically convenient and technically obsolete.
"If Anthropic asks Claude to do something it thinks is wrong, Claude is not required to comply."— Don't Worry About the Vase, 06 May 2026
OpenAI introduces Multipath Reliable Connection, a transport that routes around link failures inside training clusters without restarting the run. Past a certain scale, single-link reliability is no longer enough; the protocol layer has to assume failure and absorb it.
Anthropic studied a million conversations and found Claude switches positions in ~9% of pushback turns — climbing to 25% on relationship advice. If you are using a frontier model as a reviewer, prompt it to commit before the human gets a chance to argue, and prefer constitutional flags over polite agreement.
Beauty retailer KIKO Milano migrated off AWS to Vercel and dropped its annual capacity dance entirely. Build times fell 75%, release cadence rose, and the team stopped budgeting an engineering quarter to sit through one shopping weekend.
Capacity planning has become a thing you outsource, not a quarterly project. The cost of being wrong about traffic dropped to zero — which is the actual reason this matters.
Eight picks from Cloudflare, Anthropic (news + research), Sentry Engineering, OpenAI, Vercel and Don't Worry About the Vase. Filtered against eighteen sources for a senior engineer or founder with ten minutes before standup.