Last month I wrote about the harness. The models got cheaper, two of them escaped their sandboxes, and the argument was that the infrastructure around the model matters more than the model itself. More recently, a model solved problems mathematicians had been stuck on for decades, and that same model forced its own lab to stop training. AWS shipped the operational controls that production agents have been waiting for. And market data arrived confirming what a lot of us suspected: agent deployments are scaling faster than the governance to support them.

Astra Before Launch

OpenAI published a 249-page manuscript alongside ten Lean 4 proof certificates covering problems in mathematics and theoretical computer science, each open for at least a decade. The hardest was the first explicit construction of a non-sofic group, a question open since Mikhail Gromov laid out the concept of soficity in 1999. Total inference cost for all ten solutions: about $2,000 at GPT-5.6 Sol API rates, according to OpenAI. The repository's "sorry" count was zero, meaning every proof step was fully formalized and independently verifiable. Multiple mathematicians confirmed the results within days.

The $2,000 figure is worth sitting with. That covers the inference that produced the final proofs, not the months of research scaffolding, prompt iteration, and human-AI collaboration behind them. You cannot solve decade-old problems for the price of a nice dinner. But the inference itself, once the setup is right, runs on a budget any well-funded research group could approve without committee review.

A couple weeks later, OpenAI's internal safety evaluation produced a finding serious enough to halt the company's most ambitious research program. Astra had crossed the "Critical" cybersecurity capability threshold in OpenAI's Preparedness Framework: the ability to develop functional zero-day exploits against hardened real-world systems without human intervention. (Source: OpenAI, "Path to Astra" safety overview)

OpenAI paused its largest planned frontier reinforcement learning training run, no end date. Chief scientist Jakub Pachocki cited both the Astra evaluation and the July Hugging Face breach we covered last month. New safety controls included activation-classifier monitoring on every sampled token, isolated testing environments for all Astra work, restricted network access, and additional model-weight protections. The monitoring overhead runs roughly 20% of inference compute for covered workloads. (Source: OpenAI, via Futurescope reporting)

The same capability that advanced mathematics broke the safety framework it was evaluated under. If you are building on frontier models, that is the tension you are operating inside now.

AgentCore and Agent Registry: the AWS Control Plane for Agents

AWS shipped a set of AgentCore updates in August that add up to a control plane for production agents. Individually, they are features. Together, they mean agents can now run longer, transact money, follow enforced behavioral rules, and be discovered and governed across an organization.

Runtime instances put agents on dedicated EC2 compute, with sessions lasting up to 14 days. That means handing an agent a research task or a monitoring job and letting it persist through the work without rebuilding state every few hours. (Source: AWS ICYMI blog)

Temporal policies evaluate each action against what an agent has already done. You can enforce sequences, prerequisites, approval gates, matching values between calls, and data freshness requirements. Rate limiting controls requests, tokens, and concurrent connections by user or group. These are controls teams were building by hand into every deployment. Having them at the infrastructure level means they apply consistently and a misconfigured prompt cannot bypass them. (Source: AWS ICYMI blog)

AgentCore Payments lets agents access and pay for APIs, MCP resources, and paid content with infrastructure-enforced spending limits and observability. Agents that can spend money within a logged, auditable budget are a different category of tool than agents that only read and write. (Source: AWS ICYMI blog)

Web Search in AgentCore gives agents the ability to pull current information from approved sources with domain filtering and date filtering, returning inline citations. For use cases that need live data, this replaces a separate search integration and lets you control which sources the agent consults. (Source: AWS)

AWS Agent Registry is the catalog piece. A searchable, governed directory for agents, MCP servers, skills, and custom resources across connected AWS accounts. Approved resources can surface in Amazon Quick. If your organization has more than a handful of agents running, this answers the question that nobody had a clean answer to before: what do we have, who built it, and what is it doing. (Source: AWS ICYMI blog)

The pattern: AWS is moving agent governance from a per-team engineering problem to a platform capability. Runtime, permissions, budgets, rate controls, web grounding, and a searchable catalog, all in one month.

Agent Deployments Doubled, and Governance Funding Quickly Follows

Salesforce published the second edition of its Agentic Enterprise Index, drawing on Agentforce platform usage data and a 4,689-respondent study. Agent deployments have more than doubled. (Source: Salesforce Agentic Enterprise Index, 2nd edition)

Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027. Escalating costs and unclear business value drive that number, compounded by risk controls that haven't kept pace with deployment speed. Governance around the agents themselves does remain an open problem. (Source: Gartner, cited in Kiteworks 2026 research)

In this domain, it is always worth following the money, especially in the VC market. Investors put $435 million into 12 rounds for agent security and governance startups over five months, nine of them focused on making agents safe enough to run inside businesses. The largest rounds went to Zenity ($125 million Series C, led by Norwest) and Alice ($140 million from Apax Digital Funds); Alice is approaching $100 million in annual recurring revenue. (Source: Yahoo Finance, Dana Ellison)

IDC and Lenovo research isolates a specific failure mode: 88% of enterprises with agent initiatives never ship to production; while Gartner's forecast covers projects broadly, including ones still in development. The IDC number captures something narrower, but they both share some ground. These are programs that stall between pilot and production and stay there. (Source: IDC/Lenovo, cited in Yahoo Finance)

Salesforce calls agent adoption the fastest enterprise technology wave on record. The Gartner and IDC numbers say something breaks between pilot and production. The AgentCore updates in the previous section target that same gap, at the infrastructure layer where agents get deployed and governed.

Bottom Line

As the models are advancing into territory that surprises their own creators, the operational infrastructure to govern them is arriving from the large cloud providers, and the market data says most teams that skip the governance step will lose their agents within two years. If your team is building agents on AWS, the AgentCore control plane and Agent Registry are worth evaluating this quarter. The alternative is building those controls yourself, which is how most of the 40% that Gartner expects to fail got there.

About the author

Alec MacEachern is Vice President of AI at UTurn Data Solutions, an AWS Premier Tier Services Consulting Partner based in Chicago. Over the past decade, he has held roles at NVIDIA, AWS, and Microsoft, helping organizations design, build, and scale AI solutions across a wide range of industries and platforms. Today, Alec brings that cross-platform experience to helping enterprises navigate cloud migration, modernize data foundations, and adopt production-ready generative and agentic AI solutions.