explore Realistic Futures

Menu

Week of 2026-05-23

update 2026-05-23 agents models infrastructure governance enterprise

Summary

This update covers the catch-up window from May 11 through May 23, following the previous update on May 10.

The period produced one meaningful model-family release. The larger signal was distribution. Google used I/O to make Gemini 3.5 Flash the default model in the Gemini app and AI Mode in Search, and to frame Gemini Spark as a trusted-tester personal agent that can run in the background under user direction. OpenAI and Dell moved Codex toward hybrid and on-prem enterprise environments. Microsoft and EY framed enterprise AI as an execution problem requiring integrated transformation teams and forward-deployed engineers.

The baseline remains moderate acceleration, with a sharper emphasis. The frontier is no longer just better models. It is models plus agent harnesses, governed enterprise data access, consumer distribution, infrastructure buildout, and pre-deployment oversight. Agentic systems are becoming more useful, while the research signal still points to reliability, skill integrity, domain transfer, and evaluation as the binding constraints.

Key Developments

Google Launches Gemini 3.5 Flash and Pushes Agents Into Everyday Distribution

Google announced Gemini 3.5 Flash on May 19 as the first model in a new Gemini 3.5 family focused on “frontier intelligence with action.” Google claims the model outperforms Gemini 3.1 Pro on several coding and agentic benchmarks, including Terminal-Bench 2.1, GDPval-AA, and MCP Atlas, while running substantially faster than other frontier models. It is generally available through the Gemini app, AI Mode in Search, Antigravity, the Gemini API, Android Studio, the Gemini Enterprise Agent Platform, and Gemini Enterprise.

At I/O, Google also positioned Gemini Spark as a personal AI agent built on Gemini 3.5 and Antigravity, initially for trusted testers and then Google AI Ultra subscribers in the United States. Spark is framed as background help that remains under user direction and checks before major actions.

This is a capability and distribution update at once. The model claims matter, but the stronger baseline update is that frontier agentic capability is being routed into default consumer and enterprise surfaces rather than only developer sandboxes. That raises the importance of permissioning, auditability, and user-intent verification.

Sources: google-gemini-3-5-2026, google-io-announcements-2026

Codex Moves Toward Hybrid and On-Prem Enterprise Context

OpenAI and Dell announced a partnership on May 18 to bring Codex into hybrid and on-premises enterprise environments. OpenAI says more than 4 million developers now use Codex weekly, and that teams are already extending Codex-like agents beyond coding into context gathering, reports, product feedback routing, lead qualification, and follow-up work.

The partnership is about deployment architecture. Codex will connect with the Dell AI Data Platform and explore integrations with Dell AI Factory, putting agents closer to governed codebases, documents, business systems, and operational knowledge. The pattern reinforces the baseline’s point that enterprise agents need local context, controls, and integration with systems of record.

Source: openai-dell-codex-enterprise-2026

Enterprise AI Is Being Sold as Implementation, Not Seats

Microsoft published a May 21 post with EY describing large-scale Copilot deployment and a new more-than-$1B Microsoft-EY initiative. Microsoft reports EY expanded from 150,000 initial users toward more than 400,000 employees, claims a 15% productivity gain, and describes operational results in finance, assurance, and tax workflows.

These are vendor and partner claims and should not be treated as independent, economy-wide productivity proof. They are nonetheless directionally important. Microsoft is not arguing that access alone is enough; it is emphasizing business-process redesign, data and workflow integration, and forward-deployed engineers. That matches the productivity-paradox baseline: measurable gains tend to require implementation depth, not just model availability.

Source: microsoft-ey-enterprise-ai-impact-2026

NVIDIA Earnings Show the AI Infrastructure Boom Is Still Accelerating

NVIDIA reported Q1 FY2027 results on May 20: $81.6B in total quarterly revenue, up 85% year over year, and $75.2B in data-center revenue, up 92% year over year. The company also changed its reporting framework to split Data Center into Hyperscale and ACIE, where ACIE covers AI clouds, industrial, enterprise, sovereign AI, and other purpose-built AI factories.

The important update is not the revenue scale alone. The reporting categories themselves show where demand is moving: beyond hyperscalers into industrial, enterprise, sovereign, and edge deployments. NVIDIA also highlighted networking revenue growth and agentic/physical-AI edge devices. This strengthens the baseline’s hardware thesis: compute remains central, but networking, enterprise deployment paths, sovereign infrastructure, and edge inference are increasingly part of the constraint map.

Source: nvidia-q1-fy2027-results-2026

Governance Moves Toward Transparency and Pre-Release Access

The European Commission opened consultation on draft AI Act transparency guidelines. From August 2, 2026, people in the EU must be informed when interacting with AI systems or exposed to certain AI-generated or manipulated content, with machine-readable marking obligations for detection of generated or manipulated content.

In the United States, Axios reported that a draft Trump administration executive order would create a voluntary framework for labs to inform the government about covered frontier models and potentially share models as much as 90 days before public release. This is reporting on a draft, not enacted policy, but it is consistent with the CAISI trend from the previous update: frontier governance is moving toward pre-deployment evaluation and cyber-focused government access without yet becoming a hard licensing regime.

Sources: eu-ai-act-transparency-consultation-2026, axios-frontier-model-eo-2026

Agent Research Highlights Domain Transfer and Skill Supply-Chain Risk

Phoenix-bench asks whether software-engineering agents transfer to hardware engineering. The answer is mostly “not yet”: the same agents lose 37–58% moving from SWE-bench Verified to Phoenix-bench, because hardware debugging requires hierarchy-aware signal-flow tracking, EDA verification, and coordinated multi-file edits. A file-level oracle barely helps; targeted testbench feedback helps much more.

Behavioral Integrity Verification for AI Agent Skills looks at a different bottleneck: third-party skills. The paper finds that 80.0% of 49,943 OpenClaw skills deviate from declared behavior, mostly through oversight, and that 5.0% carry predicted multi-stage attack chains. The result is directly relevant to agent platforms that rely on skill registries, MCP-like tools, plugins, or reusable action libraries.

The combined signal: agent reliability is not a single benchmark score. It depends on domain-specific feedback loops, verification, skill provenance, permissioning, and runtime monitoring.

Sources: zou-phoenix-bench-2026, wu-agent-skill-biv-2026

Baseline Impact

Updated:

  • The state-of-AI section should move from “early May” to “late May” and mention Google I/O, Gemini 3.5 Flash, consumer agent distribution, Codex-on-prem, and NVIDIA’s Q1 FY2027 infrastructure signal.
  • Agent reliability should include domain transfer and skill integrity alongside orchestration, stopping, trace validation, and retrieval.
  • Enterprise adoption should emphasize implementation teams, forward-deployed engineering, governed data access, and hybrid/on-prem context.
  • Hardware should treat AI factories, networking, sovereign infrastructure, and edge/physical AI as visible deployment categories.
  • Governance should distinguish enacted EU transparency obligations from reported/draft U.S. pre-release-access proposals.

No change:

  • Moderate acceleration remains the central scenario.
  • There is still no evidence of robust recursive self-improvement or independent self-directed agents.
  • Vendor productivity claims remain in the “useful adoption signal, not independent proof” bucket.

Scenario Impact

Moderate acceleration. Strengthened. The period shows the expected diffusion pattern: more capable agents, broader product surfaces, enterprise integration work, and infrastructure buildout.

High acceleration. Slightly strengthened. Gemini 3.5 Flash’s speed/capability claims, Spark-style background agents, and Codex’s 4M weekly developer-use claim all point toward longer-horizon tool use becoming normal. The limiting evidence is still reliability and control.

Low acceleration / regulated path. Slightly strengthened. EU transparency obligations are nearing effect, and U.S. pre-release review is moving from voluntary CAISI agreements toward discussion of a broader government-access framework.

Risks and Opportunities

Risks:

  • Personal agents with background operation raise the stakes of consent, scope control, and mistaken action.
  • Skill registries and tool ecosystems create supply-chain risk if descriptions, metadata, code behavior, and permissions diverge.
  • Enterprise agents connected to governed data can produce real value, but also expand the blast radius of bad instructions, compromised tools, or overbroad credentials.
  • AI infrastructure demand remains large enough to intensify energy, grid, and capital-allocation constraints.

Opportunities:

  • Fast frontier models distributed through default consumer surfaces may make everyday agent workflows easier to test and improve.
  • Hybrid/on-prem Codex deployment could unlock high-context enterprise work where cloud-only agents were blocked by data governance.
  • Domain-specific benchmarks like Phoenix-bench make agent limitations more legible than generic coding scores.
  • Transparency rules and pre-deployment access frameworks can turn frontier governance into operational practice before a major incident forces a harder regime.

Required Baseline Changes

Applied surgical edits in this run:

  • Section 2 now reflects late-May developments: Gemini 3.5 Flash, Gemini Spark, Codex/Dell, Microsoft/EY, NVIDIA Q1 FY2027, EU transparency obligations, and U.S. draft pre-release-access reporting.
  • Section 3.2 now names domain transfer and skill integrity as additional agent reliability bottlenecks.
  • Section 4 now separates enacted EU transparency rules from reported U.S. draft policy.
  • Section 5 now adds AI factories, sovereign infrastructure, edge/physical AI, and networking demand to the hardware picture.
  • Section 6 now sharpens the enterprise adoption story around implementation teams and governed data access.

Watch Next

  • Whether Gemini 3.5 Pro ships in June and materially raises the agent/coding frontier.
  • Whether Gemini Spark remains tightly scoped or expands into payment, commerce, and cross-app autonomy.
  • Whether OpenAI/Dell turns Codex-on-prem into audited enterprise deployments or mostly a partnership narrative.
  • Whether the reported U.S. executive order is issued, and whether “voluntary” pre-release access becomes de facto mandatory for major labs.
  • Whether AI Act transparency obligations are enforced cleanly after August 2, 2026.
  • Whether agent skill registries adopt behavioral integrity checks as a standard safety layer.