Plumbing the
information age

The AGNTCon + MCPCon Europe 2026 event consolidated two major conferences (agent architectures and Model Context Protocol) at RAI Amsterdam on September 17-18, 2026. Post-event recap data indicates significant production maturity: 81% of attendees reported deploying agents in production environments, 89% adoption of MCP across the surveyed audience, and the AGENTS.md standard reaching 60,000 repositories. Key themes covered multi-agent orchestration patterns, stateless MCP deployment (following the July 2026-07-28 specification), agent identity and authorization, enterprise deployment patterns in regulated industries, and practical observability/governance strategies. The shift from experimental demos to production-grade reliability marks a watershed moment—practitioners are now focused on operational concerns: cost management, safety guardrails, multi-agent coordination without human intervention at every step, and regulatory compliance. Workshop topics included persistent agent memory, MCP gateway traffic routing, and securing multi-agent workflows across diverse tool ecosystems. The event reflected a maturing ecosystem where standardization and interoperability are no longer aspirational but critical to adoption.

On September 20, 2026, Google announced expanded AI agent development capabilities specifically targeting on-device and hybrid deployment models for Android and Java Virtual Machine (JVM) environments. This move addresses a critical limitation in current agentic AI stacks: most frameworks assume cloud-hosted inference with high latency tolerance. For network operations, security operations, and real-time IoT scenarios, on-device agents offer lower latency, reduced attack surface, and better privacy posture. The new tools allow developers to deploy specialized agents (anomaly detection, policy enforcement, traffic optimization) directly on edge devices or within existing JVM infrastructure. The hybrid model lets agents split reasoning: lightweight decisions run locally, complex reasoning offloads to cloud models only when needed. For NetOps and AIOps practitioners, this is significant: you can now build agents that detect and respond to network anomalies in milliseconds without cloud round-trips. The reduction in protocol calls between agent and backend systems also improves throughput for multi-agent orchestration scenarios where hundreds of agents coordinate across distributed infrastructure. Integration with existing Android and JVM toolchains means minimal disruption to current deployment patterns.

Anthropic CEO Dario Amodei published a weekend essay calling on frontier AI companies to slow the pace at which they improve model capabilities so risk prevention could keep up. OpenAI CEO Sam Altman agreed and posted his own warning about losing control of the future to AI and concentration of power. Those warnings cut into chipmakers and AI infrastructure shares while cybersecurity caught the offsetting flow, on the logic that a more dangerous threat environment expands security budgets. For investors, the more useful signal will come from upcoming earnings, where actual bookings and demand commentary from both companies will either validate or refute the notion that AI-driven security spending is increasing, rather than just being a convenient narrative for a single strong trading session. Validation arrives when CrowdStrike reports on November 25, 2026, and Palo Alto on December 1, 2026.

Anthropic announced that Claude now leads 26% of the company's model research and development work, completing most tasks end-to-end from high-level prompts while remaining under human supervision. About 90% of total R&D is conducted in collaboration with Claude, with the model handling large chunks of work under close human direction. The company operates approximately 30,000 agents performing engineering tasks as of August, with oversight systems in place to detect misbehavior. This represents a dramatic escalation from February 2026 when Claude contributed zero percent of R&D. For AIOps and ML infrastructure practitioners, this signals the maturation of agentic AI workflows at scale, raising critical questions about observability, governance, and safety validation when autonomous systems perform core engineering tasks. The rapid acceleration also highlights the need for robust monitoring and third-party evaluation frameworks to ensure safety during recursive improvement cycles.

Faculty, Accenture's AI division acquired in January 2025, will evaluate and red-team models, conduct alignment assessments, and test safeguards alongside Anthropic's internal teams. The arrangement is non-exclusive; Anthropic is also negotiating similar arrangements with research nonprofit METR and other third parties. This represents the first tangible implementation of Amodei's proposal for embedded third-party oversight in frontier AI development. The partnership includes comprehensive model assessment, adversarial testing, alignment verification, and safeguard effectiveness evaluation. While OpenAI CEO Sam Altman and Elon Musk expressed support for Amodei's framework, Nvidia CEO Jensen Huang opposed it, arguing such regulation is unnecessary. The initiative signals institutional recognition that independent verification of AI safety practices is becoming a governance expectation rather than optional.