Stillhouse Intel Brief — Apr 11, 2026

CoreWeave and Anthropic dominate the wire with a multi-year infrastructure deal covered across a dozen outlets. Meanwhile, real signal in the noise: Stanford proves AI harnesses matter more than models, Bluesky publishes a production post-mortem worth studying, and Databricks data shows agents are already reshaping enterprise data infrastructure. Artemis II splashes down after circling the Moon; Anthropic's Mythos model rattles Wall Street.

AI Tooling & Production Systems 5 articles

Stanford's Meta-Harness study delivers hard numbers: changing the orchestration wrapper around a fixed model yields 6x performance gains. The harness beat hand-engineered solutions and topped coding benchmarks without touching the model. Invest in scaffolding, not just model upgrades. Clouded Judgement 8
Calendar-based MLOps retraining is built on a flawed assumption. Empirical data shows models don't gradually forget -- they get shocked by distribution shifts. Event-driven retraining triggered by performance anomalies outperforms fixed schedules. Towards Data Science 7
SkipLabs argues the current crop of AI guardrails are cosmetic -- prompting tricks rather than structural safety layers. Production coding agents need architectural constraints baked into the toolchain, not bolted-on filters. The New Stack 7
Cloudflare ships EmDash: an open-source platform explicitly designed for AI agents to control websites, positioning it as "WordPress for agents." Infrastructure vendors are moving from experimentation to production-grade agent platforms. The Verge 6
NVIDIA releases AITune, an open-source toolkit that auto-benchmarks inference backends for PyTorch models and picks the fastest one. Removes a manual optimization step from production AI deployment pipelines. MarkTechPost 6

Bottom Line The competitive advantage in AI systems is shifting from model selection to orchestration, tooling, and infrastructure -- the wrapper matters as much as the model inside it.

AI Infrastructure Economics 13 articles

CoreWeave signs a multi-year, multibillion-dollar compute agreement with Anthropic to power Claude. Nine of the top ten AI model providers now run on CoreWeave. Stock surged 12% on back-to-back deals with Anthropic and Meta. Sherwood News 7 / The Decoder 7 / Bloomberg 6 / WSJ 6 / CNBC 6 +5 more
Meta validates the neocloud model with a $10B+ commitment to CoreWeave, the largest deal yet for a specialized GPU cloud provider. General-purpose hyperscalers are losing AI workloads to purpose-built infrastructure. Seeking Alpha 6
Anthropic launches Claude Managed Agents -- composable APIs for production agent deployment -- targeting the data center bottleneck where enterprises struggle to move from prototype to production. Data Center Knowledge 7
Databricks reports only 19% of organizations have deployed AI agents, yet those agents already create 97% of new databases. Multi-agent systems grew 327% in four months. The adoption curve is a step function, not a gradient. SaaStr 7

Bottom Line Specialized GPU clouds are winning the AI infrastructure war -- CoreWeave now hosts 9 of the top 10 model providers, and the hyperscaler moat is narrower than the market assumed.

Reliability & Platform Engineering 3 articles

Bluesky publishes a detailed post-mortem of their April 2026 outage -- real failure modes, recovery processes, and architectural lessons from operating a distributed social platform at scale. Concrete material for SLO design and incident response playbooks. pckt.blog 8
A technical presentation on the race to zero latency: decoupling business logic from I/O, using Aeron and Disruptor patterns, and consensus protocol optimization. The separation-of-concerns pattern remains fundamental for extreme low-latency distributed systems. InfoQ 6
SiriusXM built a custom platform prioritization framework that moves beyond RICE scoring, weighting developer speed, reliability, cost, and trust. The concept of "assumptions as code" -- storing and validating decision assumptions in a central repo with AI -- could reduce misalignment across platform teams. Engineering Enablement 7 Podcast

Bottom Line Production reliability still comes down to the unglamorous work -- post-mortems that get published, latency budgets that get enforced, and prioritization frameworks that account for trust.

AI Safety & Governance 4 articles

Anthropic restricts access to its new Mythos model after it demonstrated the ability to write working software exploits. The model found vulnerabilities in every major OS and browser, triggering an urgent conversation between Wall Street, the Fed, and government officials about cybersecurity risk. Fortune 6 / Dark Reading 6
The gap between enterprise AI ambitions and operational reality is widening. Pipeline sprawl and shadow AI are creating governance nightmares, and most organizations lack the tooling to even audit what their AI systems are doing. Stack Overflow 6

Bottom Line Mythos proves offensive AI capabilities are production-ready before defensive infrastructure is -- and most enterprises can't even see what their existing AI systems are doing, let alone govern them.

Monitoring — Score 5 Noise Floor 266 articles

The bulk feed is dominated by three stories: CoreWeave/Anthropic deal coverage across 36+ Anthropic-tagged outlets (press release amplification, no new signal beyond the priority articles above), Artemis II splashdown covered by NASA, WIRED, Ars Technica, Bloomberg, Fortune, and The Verge after the historic Moon flyby, and Anthropic's Mythos model sending cybersecurity stocks into freefall with coverage spanning CBS, NPR, Financial Times, and Yahoo Finance. Elsewhere: Netflix's LLM-as-a-Judge for show synopses, Etsy migrating 425TB MySQL to Vitess, Japan betting $16B on AI chip startup Rapidus, OpenAI accusing Musk of "ambush" before $100B+ trial, and Cisco in talks to buy AI security startup Astrix for $350M. The geopolitical undercurrent: Iran tensions, Colombia-Ecuador tariffs, and inflation hitting 3.3%. Product Hunt shipped 20 new tools. xkcd drew lightning.

Full Index — All 291 Articles

AI Tooling & Production Systems