From Policy to Production: AI Goes Agentic

What happened
At a FedScoop event, State Department CIO Kelly Fletcher said the department plans to “roll out agentic AI” inside its systems and will continue embedding AI in functions like translation and summarisation. She noted that the department wants to consolidate and simplify around commodity services (devices, help desks) so employees can do their jobs without administrative friction. Fletcher sees AI agents as part of that simplification; agents extend generative AI by autonomously performing multi‑step tasks and making decisions.

Why it matters
Federal agencies have embraced generative AI (the State Department’s StateChat chatbot is widely used) and are now exploring more powerful agentic systems. The Government Accountability Office (GAO) noted that even the best performing AI agents can autonomously complete only about 30 % of software‑development tasks, underscoring the need for human oversight. Analysts predict agents could be built into a third of enterprise software by 2028. Fletcher’s remarks highlight how government modernisation will combine policy, training and with agentic AI to reduce friction while ensuring safety and trust.

What Happened
OpenAI’s research-preview “Codex‑Spark” coding model is optimised for ultra‑low‑latency Cerebras hardware, pushes over 1 000 tokens per second and is designed for real‑time edits. GitHub rolled out an Agentic Workflows preview that lets developers describe a desired outcome in Markdown and run it as a workflow via GitHub Actions; they can choose their preferred coding agent (Copilot CLI, Claude Code or Codex) with guardrails for sandboxing and permissions. Google’s Gemini CLI added an automated review feature that generates post‑implementation reports and now prompts users for API keys and other settings to minimize misconfigurations. GitHub’s Copilot testing for .NET went GA, and SD Times notes Anthropic’s latest funding round adds $30 billion, pushing the Claude maker’s valuation to about $380 billion.

Why It Matters
The Codex‑Spark preview is OpenAI’s first major model to run solely on non‑Nvidia chips, hinting at more hardware diversification ahead and giving coders low‑latency, real‑time feedback. GitHub’s agentic workflows and Google’s automated reviews bring autonomous agents directly into mainstream developer tools, but with built‑in guardrails and human oversight. Anthropic’s mega‑funding underscores investors’ appetite for enterprise‑grade agents.

What Happened 
A new Alteryx survey finds fewer than 25% of AI pilots ever reach production; only 28 % of leaders trust AI for decision-making, and 49% say high‑quality governed data is the top requirement for agentic AI. To help partners build AI-ready products, Databricks rolled out a Partner Tiering program and “Well‑Architected Framework” that covers connected products, native apps and built‑ons (including MCP servers and agentic apps). Infragistics’ App Builder AI now converts sketches or spoken prompts into working React/Angular layouts, while Planview’s Connected Work Graph uses agents to map dependencies across teams and surface bottlenecks. Qlik brought its agentic analytics experience to GA and launched an MCP server so third‑party assistants can query governed data.

Why It Matters
The Alteryx findings show that trust and data quality—not algorithms—are the biggest barriers to scaling agentic AI. Databricks and Qlik are responding by offering frameworks and MCP servers, signalling a push toward secure, interoperable agent ecosystems. Low‑code generators and dependency graphs highlight how vendors are stitching context around agents, so businesses can automate without surrendering control.

What Happened
Waymo’s sixth‑generation robotaxi, Ojai, is now fully driverless on U.S. roads; the Zeekr‑based design uses just 13 cameras, four lidars and six radars. Waymo will build tens of thousands of robotaxis and expand to new cities such as Nashville, targeting more than a million paid rides per week. Anthropic closed a $30 billion Series G round, lifting its valuation to roughly $380 billion and pushing its enterprise revenue to a $14 billion run rate. OpenAI’s GPT‑5.3‑Codex‑Spark uses Cerebras Wafer‑Scale Engine 3 chips, offering a 128 k context window and near‑instant code edits.

Why It Matters
Robotaxis are moving from small pilots to industrial scale, with simpler sensor suites and ambitious ride targets. Anthropic’s funding shows that investors still believe in mega‑valuations for frontier-model providers. And the move to Cerebras chips signals that AI heavyweights are looking beyond Nvidia, which could reshape the hardware landscape.

What Happened
Reuters reports that investors are growing anxious about AI‑driven disruption, dumping shares after Anthropic launched a legal AI plug‑in and other models came out. Software, brokerage, legal‑services and insurance stocks sold off sharply, wiping a double‑digit percentage off the S&P software & services index. Chinese startup MiniMax released open‑source M2.5 models that deliver near state‑of‑the‑art performance at a fraction of the cost. Investor Matt Shumer says AI is crossing from a helpful tool into an “autonomous worker,” while marketer Ann Handley warns that enterprises need governance and balanced adoption.

Why It Matters
The sell‑off shows markets are no longer viewing AI as a free lunch; every new plug‑in could threaten an incumbent’s business model. Cheaper open‑source models will accelerate adoption by lowering the barrier to entry. But as AI becomes a “worker,” governance and data quality rise from nice-to-haves to essentials.

What to watch
Federal agencies like the State Department are pushing ahead with agentic AI, but the technology still needs trust, data governance and oversight. Developer tooling is racing toward real‑time coding and automated workflows. Investors are betting big on companies like Anthropic even as markets wobble, and hardware diversification is gathering pace. Expect more real‑world deployments—and more scrutiny—as the agentic era evolves.