This week in agentic AI, three threads converged into a single story. OpenAI shipped its most ambitious agent platform yet, making always-on, autonomous AI agents available to millions of paying users. The NIST AI Agent Standards Initiative hit a critical deadline with no enforceable framework in place, leaving enterprises to govern themselves. And the private sector responded with a flurry of governance and infrastructure products that suggest the market is no longer waiting for regulators to catch up.
Four signals stood out. First, OpenAI DevDay 2026 introduced Dots, GPT-6.1 Sol, a Decisions API, and a Marketplace, signalling that the company believes the agent layer, not the model layer, is the product. Second, BCG's Applied AI Index found that 42% of companies expect agents to act autonomously by 2030, yet only 5% have the relevant controls in place, a gap that is both a risk and an opportunity. Third, the NIST deadline for AI agent standards arrived with no enforceable framework, even as 88% of enterprises report agent-related breaches. Fourth, RSA launched Agent ID, Cloudflare shipped Clef decision models, and DigitalOcean introduced Agent Droplets, confirming that the infrastructure and governance layers are now attracting serious investment and product velocity.
The message for buyers is clear: the tools are arriving faster than the rules. For suppliers, the winners will be those who build trust architectures, not just capability demonstrations.
1. OpenAI DevDay 2026: Dots, GPT-6.1 Sol, and the shift to the agent layer
On 30 September 2026, OpenAI held its DevDay event in San Francisco and announced more than twenty updates. The headline was Dots: persistent, always-on AI agents powered by GPT-6 Astra, each running on its own cloud computer with connections to more than 4,000 applications and integrations with Slack and Microsoft Teams. Users define boundaries for what a Dot can do autonomously, what requires approval, and what is prohibited. The agents are available to ChatGPT Pro, Business Premium and Enterprise subscribers.
Alongside Dots, OpenAI launched GPT-6.1 Sol, pitched as "near-Astra intelligence for a fifth of the price." Pricing is set at $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input at $0.10 per million tokens. The company claims the model ties Astra on DeepSWE, beats Anthropic's Opus 5.5 on AutomationBench at one-third the cost, and comes within 2.1 points of Astra on OSWorld 2.0 at roughly one-seventh the price.
Other notable launches included a Decisions API for structured, confidence-scored choices; ChatGPT Spaces, a shared workspace for human and agent collaboration; a Marketplace for third-party agent extensions; and Codex cloud environments that let developers hand off bug triage, failing builds and pull requests to agentic workflows.
Why it matters: OpenAI is betting that the agent layer, not the model layer, is where enterprise value will be captured. For buyers, this means evaluating agents as production systems with lifecycle management, not as experimental features. The always-on nature of Dots raises real questions about liability, oversight and cost attribution. For suppliers, the Marketplace and extension model create both opportunity and competitive pressure, as any team can now build on top of OpenAI's agent runtime.
2. BCG: 42% of companies expect autonomous agents by 2030. Only 5% are ready.
Boston Consulting Group published its 2026 Applied AI Index this week, and the findings frame the enterprise agent market in stark terms. The research found that 88% of organisations now use AI in at least one business function. Adoption is no longer the question. Control is.
BCG reports that 42% of companies expect their AI agents to act autonomously by 2030, making decisions without human approval. Yet only 5% have the relevant controls, monitoring and governance structures in place to support that level of autonomy safely. The gap between ambition and readiness is the defining risk in enterprise agentic AI.
Why it matters: the 37-point gap between expected autonomy and actual controls is not a technology problem. It is an organisational design problem. Buyers should treat governance as a prerequisite for scaling, not a post-deployment fix. Suppliers pitching autonomous capabilities need to pair every capability claim with a clear control story, or risk being blocked by procurement and risk committees.
3. The NIST deadline arrives with no enforceable framework. Five products ship in thirteen days.
The NIST Center for AI Standards and Innovation launched its AI Agent Standards Initiative on 17 February 2026 with three pillars: agent security, agent identity and authorisation, and open-source agent protocols. The initiative was projected to deliver agent-specific overlays in the second half of 2026. Those overlays have not materialised. According to Cloud Security Alliance analysis, finalised agent-specific standards are not expected until 2027 at the earliest.
The timing matters. Gartner's 2026 CIO Survey found that 17% of CIOs have already deployed AI agents and another 42% plan to do so within the next year. AvePoint's State of AI 2026 report, conducted with Osterman Research, found that 88.4% of enterprises experienced an AI agent breach in the past twelve months. The most common incidents were data leakage at 50.1% and manipulation by malicious or untrusted inputs at 49.6%.
Into this vacuum, vendors have moved fast. Between August and October 2026, five major governance products hit the market: SAP's AI Agent Hub for vendor-agnostic inventory; Collibra's Guardian Agents for runtime supervision; Dataiku's Agent Management platform; Island's agentic control plane; and Microsoft's Copilot Autopilot with integrated Entra identity governance. This is the market governing itself because federal guidance is not yet available.
Why it matters: buyers deploying agents today are doing so without a federal measurement framework for what "safe" or "compliant" means in an agentic context. The vendor solutions are necessary but fragmented. Each defines governance differently. Buyers should treat these tools as interim measures, not permanent compliance guarantees, and build internal audit and incident response capabilities that do not depend on any single vendor's definition of control.
4. RSA launches Agent ID as identity becomes the control plane for AI
At its Next Level: Securing Identity in the Age of AI event on 22 September 2026, RSA officially launched Agent ID, a product designed to discover, inventory and govern all AI agents operating inside an organisation, including sanctioned tools and shadow AI agents. Rollout to customers is scheduled for November 2026.
Jim Taylor, president and chief product and strategy officer at RSA, highlighted a striking case study: a mid-sized global bank believed it had no AI agents in its environment. A review found more than 4,000. Taylor said this is unlikely to be an exception. A Cloud Security Alliance survey from 2026 found that 82% of organisations have unknown AI agents running in their infrastructure. IBM's Cost of a Data Breach Report 2026 found that cyber incidents involving shadow AI tools made up 43% of the total, up from 20% in 2025.
Why it matters: identity governance is the most important layer in enterprise agent security, and it is being rebuilt from scratch for non-human actors. Buyers should assume they have more agents than they know about and start with discovery, not policy. Suppliers building agent platforms should treat identity, authentication and access logging as first-class features, not security afterthoughts.
5. Decision models and agent infrastructure go mainstream
This week brought three infrastructure developments that lower the cost and complexity of running agents at scale. Cloudflare released Clef and Clef-Flash, open-source decision models optimised to return structured choices rather than freeform text, alongside an RL fine-tuning service on Workers AI. The models are designed for agent routing, guardrails and tool selection at lower latency and cost than calling a full LLM.
Strands Labs, backed by AWS, published Strands Decider 2B, a two-billion-parameter open-source decision model designed to run locally and return confidence-scored choices in tens to low hundreds of milliseconds. The pitch is practical: offload routine, high-volume decisions from expensive LLM calls to a small, fast, local model.
DigitalOcean introduced Agent Droplets, a bundled monthly plan starting at $50 for Pro and $200 for Team, packaging managed agent runtimes, serverless inference, persistent memory and governed tool access under a single predictable subscription.
Why it matters: the infrastructure layer for agents is diversifying rapidly. Buyers now have genuine choices at every tier: frontier models for complex reasoning, decision models for fast classification, local inference for privacy-sensitive workloads, and managed platforms for teams that want predictability. The cost of running an agent is falling. The cost of governing one is not. That asymmetry should shape every procurement decision.
The Agentic Expo takeaway
This was the week the agentic AI market matured from a technology story into an infrastructure and governance story. OpenAI's Dots made always-on agents a consumer and enterprise reality. BCG quantified the control gap. NIST's missing deadline proved that regulation will lag deployment. And a wave of new products, from RSA Agent ID to Cloudflare Clef, showed that the market is building the control layer itself.
For buyers, the checklist is the same as last week, only more urgent. Before scaling agents, establish identity and access controls, build observability and incident response, audit your data foundation, and design for failure recovery. The vendors who can answer those questions credibly are the ones worth shortlisting.
For suppliers, the competitive advantage has shifted decisively. Capabilities still matter, but trust architectures, deployment reliability, transparent governance and total cost of ownership are now the primary decision criteria. The market is rewarding infrastructure, not just intelligence.
The next phase of enterprise agentic AI belongs to the builders of control, not just the builders of capability.