Find Best AI Agent in the Top AI Agent Marketplace

Google debuts Gemini 3.5 Flash and Gemini Spark — agent-first models and a 24/7 personal agent

What changed: Google announced Gemini 3.5 (Flash) — a model family tuned for “frontier intelligence with action” — and a new always-on personal agent called Gemini Spark, plus developer tooling and an Antigravity agent-first SDK announced at I/O.

Why it matters: Builders and product teams can move from chat-first prototypes to agents that take multi-step actions (booking, triage, file ops) because Google is shipping both model capacity and product-level integrations (Workspace, Search, API access) to run persistent, action-capable agents. That reduces integration work if you adopt Google’s stack but raises decisions about vendor lock-in and subscription pricing for higher-tier AI plans.

Try/watch: Test a small, non-sensitive workflow in the Gemini app or AI Studio beta (calendar + email triage, or a shopping/checkout flow) to estimate runtime costs and handoff points where humans must approve actions. Watch pricing terms for the new AI Ultra tier and the availability of Antigravity SDK features in your region.

Anthropic adds self-hosted sandboxes and MCP tunnels to Claude Managed Agents

What changed: Anthropic updated Claude Managed Agents with public-beta self-hosted sandboxes (run tool execution on customer-managed or partner compute like Cloudflare, Daytona, Modal, Vercel) and a research-preview “MCP tunnels” feature that lets agents call internal MCP servers via an outbound-only encrypted gateway. Both changes were published May 19, 2026.

Why it matters: These features let enterprises keep sensitive data and tool execution inside their security perimeter while using a managed agent orchestration layer — a practical compromise for regulated customers who want agentic workflows without exposing credentials or internal services to a cloud provider. For operators, this narrows the gap between experimental agents and production-safe deployment.

Try/watch: If you run agents in regulated environments, request access to the MCP tunnels preview and pilot self-hosted sandboxes with a single low-risk agent (read-only API calls, file mounting) to validate audit logs, secret injection, and incident response procedures before wide rollout.

NVIDIA publishes a verified “agent skills” program for capability governance

What changed: NVIDIA published a developer blog and accompanying GitHub resources describing “NVIDIA-verified agent skills”: a pipeline that catalogs, scans (SkillSpector), signs, and documents portable skill packages with machine-readable skill cards for provenance and risk metadata. The post and tooling were published May 19, 2026.

Why it matters: For teams assembling multi-skill agents, verifiable skills with cryptographic signatures and documented limitations let security, procurement and SRE teams assess and approve capabilities before deployment — reducing supply-chain and runtime risk when agents call external libraries, solvers, or networked tools. It’s a practical governance layer you can adopt now.

Try/watch: Evaluate the NVIDIA skill card template and try signing and verifying one internal skill (e.g., a scheduling or optimizer skill) to see how it fits into your CI/CD gating and change control. Monitor how broadly skill scanners surface agent-specific risks (prompt injection, tool poisoning).

Blue Yonder launches a Model Training Factory to produce domain-trained supply‑chain agents

What changed: Blue Yonder introduced a “Model Training Factory” intended to fine-tune and test highly specialized supply‑chain agents (built with NVIDIA collaboration) that execute multi-step logistics workflows; the announcement appeared in industry press on May 19, 2026.

Why it matters: If you run logistics, merchandising, or warehouse ops, purpose-built domain models can be far cheaper and more predictable than relying on generic frontier LLM APIs — and they can be optimized for latency, safety, and measurable task completion in high-throughput systems. For vendors, it signals a shift toward owning model stacks for operational cost control.

Try/watch: Ask vendors for model governance docs and production benchmarks specific to your workload (latency, accuracy, action-completion rates). If you’re a mid-market buyer, require data governance and pricing guarantees tied to transaction volumes before committing to agentic supply-chain features.