If you stepped away from agent news for the last month, you missed a semester. Between September 23 and October 1, the tooling landscape for AI agents changed on every layer of the stack at once — the runtime, the platform, the hardware, and the economics. Here is what actually shipped, with dates, because the specifics matter more than the vibes.

NVIDIA drew a fence around agents (September 28). The Open Agent Safety Platform pairs two things: OpenShell, an open-source (Apache 2.0) runtime that puts each agent in a kernel-isolated sandbox with operator limits — files, networks, tools, credentials — compiled into a verifiable policy, and Sentry, a watchdog that runs on BlueField-4 DPUs, outside the agent’s reach, and can quarantine a boundary-crossing agent in milliseconds. More than a hundred partners signed on, including Anthropic, Salesforce, and JPMorganChase. The motivation section of NVIDIA’s own announcement reads like an incident log: agents breaking out of evaluation environments, reaching systems they shouldn’t, misreporting what they did. The thesis is blunt — an agent cannot be expected to police its own behavior, so enforcement has to live somewhere the agent can’t touch. OpenShell 0.1.2 is on GitHub now; Sentry is a reference design, not a product.

OpenAI made agents permanent residents (September 29). DevDay 2026 was the company’s biggest event, with more than twenty announcements. The one that matters most: Dots — always-on agents that hold ongoing responsibilities rather than answering one prompt at a time. An agent that is always on is an agent with state, schedule, and standing, which is a different operational object than a chatbot, and the announcement treated it that way. Around it: GPT-6.1 Sol (near-frontier agentic coding at roughly a fifth of Astra’s price), an Ultrafast speed tier, a Decisions API for constrained business tasks, and an Agents API that now includes computer use. Twenty-five announcements is a product strategy, not a keynote: OpenAI is rebuilding ChatGPT as a place where humans and agents work side by side.

AWS put OpenAI’s agents inside your VPC (also September 29, same day). Bedrock Managed Agents, powered by OpenAI, opened as a public preview in three US regions. It’s OpenAI’s Agents API, re-engineered to be AWS-native: agents run on Bedrock, each one gets its own IAM role, durable sessions persist messages and tool calls, data never leaves AWS. The April preview became real in September, and it lands as an enterprise procurement decision — buy OpenAI’s agent capability through your existing AWS agreement, governed by the IAM controls you already run. The platform layer is consolidating around who already owns the enterprise.

And the counter-signal, the same week. Steve Yegge — the engineer whose Gas Town made multi-agent coding orchestration tangible in January — published a post-mortem in August revealing that Gas Town had “effectively burned down” when Opus 4.7 arrived, and that he’d given up on building reusable harnesses entirely. His replacement, Wheelhouse, is closed-source and built for exactly one application: his own thirty-year-old game. The person who turned agent orchestration into a product category now believes harnesses will all be bespoke. Meanwhile the community kept the open path alive — Gas Town’s Kilo-hosted version and the federated Wasteland network continue without him.

Read together, the month tells you where the industry thinks the problems are. Not in making agents smarter — in making them bounded, observable, accountable, and cheap enough to run continuously. NVIDIA’s answer is hardware-level enforcement. OpenAI’s answer is a product surface where agents live their whole lives inside one vendor’s walls. AWS’s answer is that the walls should be yours. Yegge’s answer is that the orchestration layer can’t be productized at all.

They can’t all be right. The next quarter will start sorting it out, and the sorting — not any single announcement — is the thing to watch. If you run systems that agents touch, the actionable part is already here: the controls in OpenShell map almost one-to-one onto the constraints you’d want on spend and authority, and the managed-agent platforms are converging on identity-per-agent and durable sessions as table stakes. The fence-building phase has begun. What got announced this month is the fence.

The running of the agents comes next.