Agents Act. Systems React. Control Gets Claimed.

What happened
Forbes Tech Council outlined a “Golden Five” framework for securing agentic AI systems focusing on identity, observability, and governance as non-negotiable foundations for enterprise deployment.

Why it matters
As autonomous agents move from copilots to operators—executing workflows, calling tools, and making decisions, security risk shifts from data leakage to real-world action. The takeaway: agentic AI isn’t just another SaaS tool. It behaves more like infrastructure and needs to be treated like it.

Traditional security models (perimeter-based, human-in-the-loop) break down when agents act independently. That creates new attack surfaces: compromised identities, invisible decision chains, and unchecked automation loops.

What’s next
Expect enterprise stacks to evolve fast. Identity layers will get stricter. Observability will move from logs to real-time agent tracing. Governance will shift from policy docs to enforced runtime controls.

What happened
Google announced a $750 million commitment to accelerate partner-developed agentic AI solutions across consulting firms and software vendors. The program focuses on co-building solutions using Gemini models and Google Cloud infrastructure, embedding AI into partner products and enterprise workflows.

Why it matters
This is not just a marketing fund. This is a distribution play. Instead of owning every use case, Google is turning partners into its agent layer—spreading Gemini across CRM, ERP, and industry-specific systems.

What’s next
Expect other cloud providers to respond with their own partner bets. The investment may compress adoption cycles as private‑equity‑backed firms are required to implement Gemini across their holdings.

What Happened
Meta announced a multibillion-dollar partnership with Amazon Web Services (AWS) to deploy tens of millions of AWS Graviton5 processor cores, specifically to power its agentic AI workloads. This brings purpose-built silicon together with AWS's AI stack for Meta's autonomous agent systems at hyperscale.

Why it Matters
This is one of the clearest signals yet that agentic AI is no longer a roadmap item—it's a production infrastructure priority. Meta is simultaneously cutting 10% of its workforce to double down on AI, underscoring where the investment is flowing.

What’s Next
Expect other hyperscalers to follow with custom silicon strategies for agentic workloads, and Meta to accelerate agentic deployment on this expanded compute base.

What happened
Delivery Hero unveiled Herogen, an autonomous software‑delivery agent built on multiple large‑language models. Herogen writes, tests and iterates on code from natural‑language prompts; a “council of agents” reviews each proposal before a human approves it. Rolled out to only 18 % of developers, Herogen already handles 9 % of all code‑change requests and outputs the equivalent of 130 senior engineers. It merges over 100 pull requests per day with an 85 % success rate, freeing roughly 250,000 hours of manual coding annually.

Why it matters
Herogen moves beyond copilots by autonomously shipping code. Delivery Hero says engineers can delegate mundane tasks and focus on architecture and innovation. The tool reframes automation as job evolution rather than elimination.

What’s next
The company plans to have Herogen handle 20 % of code changes by year‑end. Its success could pressure other large tech teams to deploy similar agentic coding systems.

What Happened
OpenAI launched GPT-5.5, optimized for coding and research workflows with more autonomous capabilities. Anthropic released Claude Opus 4.7, emphasizing literal instruction-following and improved output quality. China's DeepSeek previewed V4, featuring a 1-million-token context window and hybrid attention architecture with versions tuned for both cost and performance.

Why it Matters
Simultaneous frontier launches compress the competitive timeline and signal that model capability improvements are now arriving in waves, not slow rollouts. DeepSeek V4's 1M context window in particular unlocks longer and more complex agentic task chains.

What’s Next
Benchmark wars will heat up immediately, and enterprise buyers will face a choice between three genuinely competitive frontier models—a dynamic that will drive pricing pressure and accelerate adoption.

What Happened
DeepSeek has adapted V4 to run natively on Huawei's Ascend AI chips, a direct response to US export controls cutting off access to NVIDIA GPUs. China's top AI labs—Baidu, Alibaba, Tencent, ByteDance, and DeepSeek—are all accelerating domestic chip integration.

Why it Matters
If DeepSeek V4 achieves competitive performance on Huawei silicon, US export controls lose a significant portion of their leverage. This pivot is the clearest signal yet that China's AI ecosystem is building a parallel hardware stack, not just waiting out the restrictions.

What’s Next
Cost curves on Huawei-based clusters could drop sharply as domestic supply scales, potentially enabling China to train and serve frontier models at lower cost than US-dependent competitors.

What happened
Autonomous Resource Corporation (ARC) and Oak Ridge National Laboratory announced a memorandum of understanding to create the Exascale Foundry, a public‑private partnership to accelerate AI‑enabled defense manufacturing. The plan marries ORNL’s high‑performance computing and advanced manufacturing facilities with ARC’s ARCNet distributed manufacturing platform. The partnership will deploy seven production nodes connected via ARCNet and integrate ORNL’s Peregrine AI, which has analyzed 1.9 million additive‑manufacturing layers. It aims to compress production timelines for mission‑critical components from years to months and focuses first on nickel superalloy turbine parts for autonomous air‑vehicle engines.

Why it matters
U.S. defense supply chains struggle to scale additive‑manufacturing capacity. By combining ORNL’s exascale compute and ARC’s autonomous manufacturing grid, the Exascale Foundry could deliver qualified parts quickly and at volume. The initiative supports the DOE’s Genesis Mission to rebuild critical manufacturing infrastructure.

What’s next
ARC will deploy production nodes and ORNL will contribute HPC and materials‑science expertise. If successful, the model could expand beyond defense to other industries seeking federated, AI‑driven manufacturing.

What Happened
AI drone maker Skydio unveiled a $3.5 billion plan to expand US-based manufacturing, onshoring critical components and reducing reliance on overseas supply chains. The company described the initiative as a national security imperative that will create thousands of American jobs.

Why it Matters
Drone technology is now explicitly framed as physical AI infrastructure for defense, public safety, and critical infrastructure—and the US government is backing domestic production as a geopolitical necessity. This mirrors the chip onshoring playbook applied to robotics.

What’s Next
Expect increased geopolitical competition in AI-enabled drone systems, with other defense-adjacent robotics companies likely to follow Skydio's onshoring strategy.

Agentic AI is crossing the threshold from tool to infrastructure where security, distribution, and compute are no longer support layers, but the system itself. As agents execute in the real world, the winners will be those who control identity, ecosystem reach, and the silicon underneath it all.

Build your own “mini agent stack” in one hour. Use a frontier model like OpenAI GPT-5.5, Anthropic Claude Opus 4.7, or DeepSeek V4 to create a simple agent that writes code or automates a task. Then layer in a second “review agent” (inspired by Delivery Hero’s Herogen) to validate outputs before execution.

Next, simulate the real shift: add basic guardrails and define what the agent can and cannot do, log its actions, and track decisions step-by-step. You’ll quickly see the core insight from today’s stories: once agents start acting, you don’t just need intelligence, you need identity, oversight, and control.