The Agentic Turn: How AI Is Learning to Act, Not Just Answer

If you've been paying attention to AI news this month, you may have noticed something interesting: the conversation has quietly shifted away from what AI can say and toward what AI can do — and, crucially, how much it's allowed to do.

Two developments from the past few weeks crystallize this shift better than any announcement about raw model capability. Together, they tell a coherent story about where the industry is actually heading.

Decision Models: When AI Stops Writing and Just Picks

For most of the past few years, AI meant a model that reads your prompt and writes something back. That paradigm is still dominant — but a new category has emerged this fall that deliberately abandons text generation altogether.

A decision model is a small AI model that answers typed questions about some input — yes/no, pick one of these, score it 1 to 4 — and returns each answer with a probability, instead of writing text.

If that sounds simple, that's the point. These models do not generate paragraphs. Instead, they return structured choices, scores, or probabilities drawn from options defined by the developer.

TypeSafe's Jev kicked off the category on September 15, 2026. By October 1, Cloudflare had shipped Clef and Clef-flash, Amazon's Strands Labs had shipped Strands Decider 2B, and OpenAI had announced a Decisions API. The speed of adoption has been striking: Vercel's October AI Gateway Production Index reports TypeSafe's Jev decision model as the gateway's fastest-adopted model ever, reaching 35% of paid teams in two weeks.

Why does this matter for AI creators? Think about the agentic workflows many of us are already building — pipelines where an AI needs to route a request, flag a piece of content, classify an image, or decide whether to escalate a task. Today, most of us either write elaborate prompts and parse the text output, or set up a full LLM call just to get a yes or a no. Decision models offer a leaner, faster, more reliable alternative for those specific branching moments.

Because the output space is constrained, software can consume the result without extracting an answer from generated prose. Several models also target calibrated probabilities, although calibration must be verified on the intended task. That last caveat matters: decision models are not smarter than LLMs — a reasoning LLM still tops the accuracy charts. They're not a replacement for nuanced judgment. They're a purpose-built tool for structured, high-volume decision points inside a larger workflow.

The options already span a range of deployment styles. Models such as Jev, OpenAI Decisions API, Microsoft-Decision-1, Cloudflare Clef, Perplexity Decider, Liquid AI D1, and others differ in cost, latency, modalities, deployment, licensing, and customization. For AI creators building pipelines, that's a real menu to work from — not just a theoretical concept.

The Identity Problem: Who Vouches for the Agent?

At the same time decision models are maturing, a separate and equally important question has bubbled up: if AI agents are going to act in the world — booking things, filing things, moving data between systems — how do we know what each agent is authorized to do?

Right now, most enterprises running AI agents face a structural gap. Zero-trust architecture — the security model that requires every entity to prove its identity before accessing any resource — was built over the past 15 years for humans and devices. AI agents have sat outside that framework: they can be instantiated at scale, inherit shared API keys or service credentials, and take actions across system boundaries without any discrete cryptographic identity that an audit trail can pin down.

On October 15, ServiceNow took a direct shot at this problem. The company announced AI Passport, a system that uses cryptographically signed credentials to gate what AI agents can do at runtime. Each agent — whether it's ServiceNow's own or a third-party agent — carries a credential that states its identity, permitted actions, accessible systems, and accountable owner. Agents check that credential as the work runs.

The enforcement model is worth noting. The system allows AI agents to operate within defined parameters and triggers graduated responses when agents drift from their authorized scope, including termination capabilities. The urgency behind this is backed by data: 100% of banks use AI, but only 24% have governance frameworks.

The company is emphasizing that AI Passport is built on open specifications — intended to allow any governance product, identity provider, or enforcement point to issue and verify credentials, rather than locking users into a proprietary silo. A few days earlier, on October 6, a parallel effort had also surfaced at the protocol level: Sierra and Meta published Personal Agent Protocol v0.1, an open OAuth-based standard that lets a user's AI agent authenticate with businesses and declare its scope on arrival.

For AI creators, this isn't just an enterprise IT story. Anyone building agents that interact with external services — APIs, platforms, payment systems, creative tools — is going to need to think about identity and scope sooner than they might expect. The infrastructure is being built now, while many of us are still prototyping.

What It All Means

Looked at side by side, decision models and agent credentialing are solving the same underlying problem from different angles: how do you take AI from something that talks to something that acts reliably at scale?

Decision models handle the routing and classification moments inside a workflow — fast, structured, auditable. Agent credentials handle the authorization layer — who is this agent, what can it touch, and who's accountable? Neither of these is glamorous frontier-model news. But they're exactly the kind of infrastructure that makes agentic AI usable in production rather than impressive in demos.

For creators building on AI platforms, the takeaway is practical. When you design a workflow today, think in terms of the moments where a full generative model is overkill versus where you need a structured choice. And as you build agents that reach outside your own tools, start thinking early about scope and permissions — not as an afterthought, but as part of the design.

The era of AI as a text box you type into is giving way to something more like a team of specialized components. The interesting creative and professional challenge is learning to compose those components well.

The Agentic Turn: How AI Is Learning to Act, Not Just Answer

Decision Models: When AI Stops Writing and Just Picks

The Identity Problem: Who Vouches for the Agent?

What It All Means

Sources