Rewriting the Playbook for Agentic AI

Many organizations rush to prototype with AI agents, experimenting with small tasks. But these experiments often stay stuck as disconnected demos, disjointed from real project flow, lacking the context, controls, and documentation needed to produce meaningful outcomes. The leap from prototype to production is a chasm, not a step. Crossing it doesn’t require bravery, it requires an adoption playbook that turns risk into repeatable practice.

This is a subject we at frog recently explored in ‘AI Agents in Action: Foundations for Evaluation and Governance’, a joint white paper produced by the World Economic Forum in collaboration with Capgemini.

AI agents are being pitched as productivity accelerators. They can plan, decide, and act across systems without waiting for a human to supervise every step. With their ability to take action, use tools, and operate under uncertainty, AI agents are unlike any preceding innovation. They work at a pace and scale that breaks traditional oversight, so governance can’t be an afterthought.

That’s why we advocate progressive governance: practical “speedbumps” embedded in workflows that include clear guardrails, verification points, escalation rules, and human approvals at the moments that matter. The goal is not to slow agents down. It’s to let them run safely inside organizational boundaries.

Business leaders want to lead AI agent adoption and gain a competitive advantage. But leaders cannot treat deployment like a race because agent failure modes can escalate fast: brand damage, financial loss, security exposure, and in some settings, physical impact. A new playbook is needed that is fundamentally different from traditional software rollout. It requires an operating model of progressive governance that is one part IT, one part risk, one part people leadership that continuously improves performance while keeping agents set with the minimum authority required to do the job.

Real transformation won’t come from clever agent demos. It comes from giving AI agents real context: an understanding of how work gets done, access to the right trusted data (and no more than necessary), and continuous feedback from human experts. Only then can agents move from tools to true collaborators embedded in the fabric of the organization.

Making that shift requires a new mindset around governance: progressive, shared, and operational. This isn’t something a single team or function can “own.” Every role has a stake in how agents are designed, deployed, and held accountable. The real question isn’t “can we build agents?” It’s “are we ready to run them?”

Why AI agents change the risk equation

AI agents are not simply more capable chatbots. Using a combination of models, tools, memory, and systems access, AI agents can plan, decide, and act toward goals. But it is their ability to do all this with minimal supervision that makes them so unique.

In customer support, an AI agent can autonomously resolve cases end‑to‑end. In marketing, an AI agent can turn data into hyper‑personalized content. In hospitality, an AI agent can summarize staffing gaps and predict seasonal booking based on data. In scientific research, an AI agent can deconstruct a question, search multiple sources, and synthesize recommendations without step‑by‑step guidance. And that autonomy is exactly what makes the stakes higher.

Figure 1: Software architecture of an AI agent
Source: AI Agents in Action: Foundations for Evaluation and Governance | World Economic Forum

This shift from “model” to “agent” matters because it changes the risk profile. Agents can misuse tools, drift from intended goals, or interact unpredictably with other agents. They operate in environments that are often dynamic, incomplete, and uncertain.
That’s the trade: agents move closer to real work, and real work is messy. You get speed and coverage, but also drift, edge‑cases, and compounding errors if oversight isn’t designed in. And yet, adoption is accelerating. Many organizations plan to integrate AI agents within the next one to three years. Most efforts today are still pilots which is precisely why now is the moment to set the rules before the rules are set by accidents.

The mistake leaders risk making

The most common failure mode is not technical; it is organizational. Many companies approach AI agents the way they approached previous automation: deploy quickly, restrict access, and assume governance can be solidified later. But agents are not static systems. They behave differently depending on context, autonomy, authority, and environment. Governance models designed for traditional software (e.g., focused on uptime and access control) are no longer sufficient.

A useful analogy is hiring. When a person joins an organization, they are not given full authority on day one. They are assigned a role, trained, observed, and gradually entrusted with more responsibility. Performance is evaluated continuously, not just at onboarding. AI agents deserve the same discipline: scope, training, evaluation, and progressive trust.

From isolated agents to agentic ecosystems

Much of what has been discussed concerns the implementation of individual agents for specific purposes: customer service, finance, research, etc. For many organizations, that is where exploration stops. But the real gains emerge when you move from the singular to the plural: multi‑agent ecosystems that begin to connect work across the organization.

As agents begin to collaborate, transact, and negotiate across teams, vendors, and organizations, new possibilities open up alongside new considerations. These systems must navigate shared goals, dependencies, and handoffs in environments where no single team has full visibility or control.

Standardizing interaction doesn’t guarantee safety, but it does create a foundation for scale. Without boundaries on authority, verification of intent, and human‑level governance, ecosystems can propagate errors as efficiently as they propagate value. With those elements in place, they can also unlock speed, resilience, and entirely new ways of working.

Treat agent‑to‑agent connections as an organizational interface, not just a technical integration, by putting explicit constraints, monitoring, and accountable ownership in place. Leaders who thoughtfully approach classification, evaluation, and governance are far better positioned to harness the opportunity of agentic ecosystems while staying in control. In doing so, they create the conditions for complexity to become an advantage, not a liability.