• Why Higher Ed's Pandemic Playbook Is the Blueprint for AI Infrastructure — Higher ed IT leaders are drawing direct parallels between pandemic-era infrastructure modernization and the current AI rollout, arguing that institutions which invested in hybrid cloud flexibility before 2020 are the same ones best positioned for AI now. 🔗 Graph: Higher Ed AI, AI Strategy, Enterprise Monitoring, AI Governance 📅 Published: 2026-07-16 📰 https://edtechmagazine.com/higher/article/2026/07/why-higher-eds-pandemic-playbook-blueprint-ai-infrastructure 📌 Key takeaways: • AI workloads should be integrated into existing infrastructure platforms rather than treated as specialized, isolated projects requiring niche hardware and dedicated staff — the same lesson learned during the pandemic pivot to remote learning. • Fragmented AI governance creates real risks: when AI runs outside central IT, specialized hardware creates single points of failure, data duplication strains budgets, and faculty turn to shadow IT for quick access. • Regulatory constraints (FERPA, HIPAA, grant-funded clearance requirements) make integrated governance non-optional for universities — AI systems must handle data classification boundaries that corporate environments rarely face. • Nutanix positions AI as "just another workload" on existing hyperconverged infrastructure, reducing the need for siloed GPU environments — a model that aligns with how UCSD already runs TritonAI on SDSC infrastructure. • For Brett: this validates the TritonAI architecture of running AI as an integrated workload on existing campus infrastructure rather than building a separate AI stack, and reinforces the governance framework already in place through the AI Cabinet.

• Microsoft Moving to Internally Developed AI Models in Office Apps — Microsoft has begun routing tens of thousands of weekly Excel and Outlook prompts to its own MAI models instead of OpenAI and Anthropic, signaling a shift from frontier model pursuit to deployment economics as the next enterprise AI battleground. 🔗 Graph: Microsoft, Azure OpenAI, Model Agnosticism, LLM Gateway, AI Strategy 📅 Published: 2026-07-14 📰 https://campustechnology.com/articles/2026/07/14/microsoft-moving-to-internally-developed-ai-models-in-office-apps.aspx 📌 Key takeaways: • Microsoft is deploying its in-house MAI model family (introduced at Build 2026) for select Excel and Outlook workloads, with MAI-Code-1 reportedly delivering coding performance comparable to Anthropic's Opus 4.6 at lower cost. • The move illustrates a broader architectural shift: enterprises are assembling portfolios of models optimized for different tasks rather than relying on a single frontier model, using multi-model routing to match workload to the most cost-effective option. • Satya Nadella's framing: long-term AI leadership depends on deployment economics and infrastructure, not just model capability — a thesis that directly validates LiteLLM's model-agnostic routing approach. • For Brett: this is strong external validation of the TritonAI gateway architecture. Microsoft is doing at scale what LiteLLM Enterprise does for UCSD — routing to the most cost-effective model per task. The multi-model portfolio approach is exactly the strategy Brett has been building.

• What building Shippy taught us about building agents — AI2's Skylight team shares the architecture behind Shippy, a maritime AI agent for real-time ocean monitoring, detailing how they built a reliable agent system for high-stakes operational decisions where wrong answers have real-world consequences. 🔗 Graph: Agentic AI, Model Context Protocol, AI Governance, AI Strategy 📅 Published: 2026-07-15 📰 https://huggingface.co/blog/allenai/shippy-tech-blog 📌 Key takeaways: • Shippy's architecture separates "soul" (system prompt and behavioral boundaries), "skills" (markdown-based tool definitions following the same agent-skills spec used by Claude Code and Codex), and "config" (runtime settings including model and harness choice) — a clean separation that makes agents versioned, testable, and swappable. • The agent uses deterministic tools (API queries, boundary lookups, map generation) wrapped in a nondeterministic LLM reasoning layer, with every response showing its work: data sources, query timestamps, and deep links for analyst verification. • Shippy runs on OpenClaw, an open-source agent framework, with Claude Opus 4.6 as the base model — making the harness and model independently swappable via config changes. • Evaluation focuses on agent reliability, not model benchmarks: the team tests against live Skylight data (continuously updated satellite and vessel signals) rather than static snapshots, ensuring the agent degrades gracefully when data is stale or ambiguous. • For Brett: the soul/skills/config separation and the emphasis on verifiable agent outputs mirrors the TritonAI Harness architecture. The use of standardized skill specs compatible with Claude Code and Codex is directly relevant to the UCSD Skills Library work.

• MemoHarness: Agent Harnesses That Learn from Experience — Researchers introduce a framework for adaptive agent harness optimization, decomposing the harness into six editable control dimensions and learning from past executions to improve future agent performance without test-time labels or feedback. 🔗 Graph: Agentic AI, AI Governance, AI Strategy, Claude Code 📅 Published: 2026-07-14 📰 https://arxiv.org/abs/2607.14159 📌 Key takeaways: • An "agent harness" — the external control layer managing context, tools, orchestration, memory, decoding, and output handling — has more impact on agent behavior than the base model choice, yet most deployed agents use a single static harness configuration for all cases. • MemoHarness stores per-case diagnoses and distilled global patterns in a dual-layer experience bank, adapting the harness to each new task using retrieved experience — improving performance across shell-agent, code-generation, and analytical-reasoning benchmarks. • The framework remains cost-competitive when retrieved experience is cacheable, suggesting that experience reuse is a practical scaling strategy for production agent deployments. • For Brett: this directly informs the TritonAI Harness architecture. The concept of an adaptive harness that learns from execution history aligns with the harness evolution Brett's team is building, and the six editable control dimensions provide a useful decomposition for thinking about harness design.

• "AI + Education" Action Plan — CSET publishes a translation of the Chinese government's comprehensive plan for integrating AI across its education system, from using AI to plan school construction and university majors to deploying AI agents for scientific research and classroom surveillance. 🔗 Graph: Higher Ed AI, AI Governance, AI Strategy, AI Adoption 📅 Published: 2026-07-13 📰 https://cset.georgetown.edu/publication/china-ai-plus-education-action-plan 📌 Key takeaways: • China's State Council plan mandates AI integration across all education levels — using AI for school site planning, university major selection, scientific research acceleration, and even classroom surveillance of teachers and students. • The plan calls for AI agents to accelerate scientific research and for AI-powered educational resource allocation, positioning state-directed AI adoption as a competitive advantage in education policy. • This is a whole-of-government approach: the plan references Xi Jinping's education strategy and ties AI education deployment to national security and economic competitiveness goals through 2035. • For Brett: this is the geopolitical counterweight to US higher ed's decentralized AI adoption. While UCSD builds TritonAI incrementally with governance and faculty buy-in, China is mandating top-down AI integration across its entire education system. Useful context for cabinet-level conversations about AI strategy and competitiveness.

• Only 26% of enterprises say AI governance keeps pace with deployment, Smarsh study finds — A 2026 enterprise AI trends study reveals a 29-point gap between AI deployment (55% of enterprises) and governance readiness (26%), with shadow AI detection capabilities at just 30%. 🔗 Graph: AI Governance, AI Compliance & Governance, AI Security, AI Adoption 📅 Published: 2026-07-16 📰 https://www.marketscale.com/industries/software-and-technology/only-26-of-enterprises-say-ai-governance-keeps-pace-with-deployment-smarsh-study-finds 📌 Key takeaways: • While 55% of enterprises are actively deploying AI, only 26% report governance frameworks fully aligned with their deployment pace — a gap that translates directly into compliance exposure, especially in regulated industries. • Shadow AI is the compounding problem: only 30% of organizations can detect AI tools employees use outside approved workflows, creating supervision and recordkeeping blind spots that regulators are increasingly scrutinizing. • Enterprises are shifting budget in response — 62% investing in AI/ML capabilities, 53% in data quality, and 51% in modernizing archives — treating communications data as a strategic AI foundation rather than just a compliance liability. • For Brett: UCSD's proactive governance through the AI Cabinet and TritonAI's centralized gateway architecture put the institution ahead of the 26% baseline. The shadow AI detection gap is a concrete argument for expanding the Developer API Program to capture currently unmanaged AI usage across campus.

💡 Signal: This week's stories converge on a single theme: the AI infrastructure battle is shifting from model capability to deployment economics and governance. Microsoft's internal model routing, AI2's agent architecture lessons, and the 26% governance gap all point to the same conclusion — the institutions that win at AI will be those that treat it as an integrated, governed workload rather than a frontier science experiment. Brett's TritonAI architecture (LiteLLM gateway, centralized governance, skills-based agents) is already built on this thesis.