Are AI agents for business the next step in automation?
AI agents are the “third wave” of artificial intelligence, evolving from predictive AI and generative AI (GenAI). These AI-powered programs don’t just chat. They act. Specifically, AI agents can autonomously sense their environment, make decisions, and execute tasks.
For business leaders and middle managers, AI agents are akin to new team members. They can help teams accomplish more work with fewer resources and in less time. For example, they can assist with administrative tasks, improve team collaboration, and enhance customer experiences and personalization.
The field of AI development that allows AI agents to be built is known as agentic AI. Using standards like the Model Context Protocol (MCP), AI agents can connect to the external tools and data sources they need to complete their actions.
But AI agents can’t exist in a vacuum. Even though they’re autonomous in nature, they require leadership, management, and supervision — just like any other team member. And just like any other team member, achieving desired business results requires integrating them into existing business processes and workflows so they support the organization’s strategic objectives instead of distracting from them.
How can managers direct and supervise autonomous AI agents?
Effective AI agent management using a human-in-the-loop approach enables responsible AI. Apply the following strategies to keep agents focused on their intended outputs:
Use quality data
Data that AI agents use must be high quality. If agents pull from outdated, inaccurate information, they’ll incorrectly think that’s the standard. A spreadsheet or list of data in different formats — like dates or financial sets — can also confuse agents.
The same mindset applies to agent iteration. Agents will consider the sources they’re given — without parameters, they may scour the entire Internet to find the right solution, potentially leading to misinformation. Companies can prevent this by limiting what agents can access.
Set clear rules, goals, and boundaries
An AI agent is like a new employee: If it never learns company rules and goals, it could easily violate them.
System prompts help contain agents by aligning outputs with desired outcomes. A system prompt defines agent behavior through contexts, such as the agent’s persona, a task with instructions on what to accomplish, and guidelines the AI must follow. System prompts might also include elements like focus areas, tone, and example outputs.
Make these instructions as clear and specific as possible. Contradictions — for example, asking for a “strong, authoritative tone with a friendly approach” — may confuse an agent. Similarly, telling an agent it’s a “thought leader” in a topic is broad and leaves too much open to interpretation.
Begin with simple tasks
AI agents for business are best suited to handling straightforward tasks. More complicated tasks typically require multiple agents to serve as checks and support for each other. An agent must also be able to perform what it’s asked to do. For example, asking for a budget forecast is fruitless if the agent can’t access financial data.
Incorporate rate limiting
Rate limiting can help prevent AI agents from performing a bad action too frequently. For example, rate limiting would stop an agent from logging in to a service repeatedly or accessing sensitive user data.
Develop strategic authentication and authorization
Because agents resemble human employees, they must follow proper authentication and authorization processes. Verify that agents are who they claim to be by using client credentials with secure cryptographic keys.
Agents rarely need access to everything. Keep them focused with limited access tailored to their specific purpose, adjusting permissions for resources, sensitive data, and recent agent actions. For example, you might give an agent read permissions but not write permissions so it can’t modify a file.
Developers can configure authentication and authorization for their MCP server to define scopes and permissions and enforce them so agents can only invoke what’s permitted. An MCP client can also build agents that prompt the end user to grant access to another service; if the request looks strange or unnecessary, the human user can restrict that access.
Likewise, agents rarely need to be “always on.” It can therefore be advantageous to program them to have time limits. If they’re being used to craft social copy for an upcoming announcement, for example, they only require a small window of time to digest relevant news, identify tone in previous content, and craft the proper post. Limiting access time reduces potential cybersecurity risks.
Consider the OWASP risk of excessive agency
The Open Web Application Security Project (OWASP) is an international nonprofit focused on improving web application security. One risk from the OWASP Top 10 report for large language models (LLMs) is excessive agency.
LLM-based agents often have some independent agency, responding to prompts by taking action. However, excessive agency may cause issues. For example, the agent might delete files, share sensitive information, or accidentally break other operational processes.
To prevent excessive agency, developers can limit the functionality, permissions, and autonomy of plugins and tools. For instance, they can use rate limiting for items like message sending or chat completions. Authorization and authentication with OAuth is another sound strategy against excessive agency, as is requiring human authorization for specific actions in order to mitigate negative impacts.
Think from a UX perspective
Employees will have varying knowledge about the inner workings of AI agents. They may only see the end result. If an agent is meant to assist all team members, it must be simple to use.
Consider what AI agents can already do, then remove any potential derailing distractions or extra steps. Regularly reviewing rules and guidelines can help uncover extraneous areas in an agent’s system.
What’s the best way to integrate AI agents into existing workflows?
AI agent management is one piece of the puzzle, but successfully implementing AI agents for business also requires embedding them into existing workflows in natural, comfortable ways. These tactics can help facilitate seamless integration and ongoing operation:
Conduct workflow audits to uncover potential improvements or bottlenecks, highlighting tasks and processes AI can optimize or automate.
- Ask AI providers with API-based integrations to help you examine where AI can fit into existing systems.
Start small and gradually increase scale, offer employees training and resources on how to use agents, and encourage team communication.
Include a human-in-the-loop workflow with state persistence and human evaluation of an agent’s decision-making process.
Remember that AI agents are teammates, not replacements.
Successful AI agent management: How do you measure it?
If you want to grow the use and impact of AI agents in your organization, performance tracking must be a core component of your AI agent management strategy. Possible metrics include:
Completion rate and the number of steps it takes an agent to complete a given task
Accuracy, including whether the agent followed instructions, whether it achieved expected results, and whether a human had to step in to make adjustments
Time saved by using the agent, particularly for administrative tasks
Employees are also a great source of feedback. Because they use agents regularly and have frontline experiences, they can share successes and potential challenges, offer benchmarks, and provide examples for others to help them get comfortable using and managing AI agents.
Managers are accustomed to monitoring human employees and giving them regular performance reviews. They must similarly monitor agents by consistently performing evaluations, tracing, and failure analysis to uncover common mistakes and help “debug” inaccurate or inefficient agents.
How can Cloudflare help with AI agents?
The Cloudflare developer platform has tools to build and deploy AI-powered agents and keep them running smoothly. Especially helpful are:
Workers AI: Workers AI provides globally distributed compute power for running AI agents close to where decisions happen, enabling real-time responsiveness and edge-based autonomy.
Cloudflare containers / sandboxes: Cloudflare containers and sandboxes are secure execution environments that can isolate and manage AI agents, ensuring governance and policy enforcement. They’re ideal for managing agent lifecycles in production.
Agents SDK: Cloudflare empowers scalable, secure AI agent deployment through Agents SDK. These agents come with built-in state management, communicate with clients in real time, and use desired AI models and data to help businesses operationalize AI responsibly and securely.
MCP servers: Remote MCP servers can open up LLMs and agents to a larger audience of users by fitting into regular sign-in and authorization flows. Rather than sharing passwords, remote MCP servers require authentication and authorization in the form of OAuth.
MCP Server Portals: Instead of distributing dozens of individual server endpoints, developers can register their servers with Cloudflare and provide a single, unified portal endpoint for users to configure in their MCP client. This is like a centralized front door for MCP servers.
Running AI agents on Cloudflare can also help control costs: Cloudflare charges only for compute time (when the agent is thinking) and not wall time (when the agent is waiting).
Ready to see how AI agents can save time and enhance productivity? Build your first agent here.
FAQs
What is the primary purpose of an AI agent for a business?
AI agents are designed to act autonomously, helping business teams complete more work with fewer resources and in less time. They can assist with administrative tasks, improve team collaboration, and enhance customer experiences.
How can managers effectively manage AI agents?
To effectively manage AI agents, managers should use high-quality data and set clear rules, goals, and boundaries through system prompts. Start with simple tasks, use rate limiting, and develop strategic authentication and authorization processes.
What are the key strategies for integrating AI agents into existing business workflows?
To successfully integrate AI agents, conduct workflow audits, start with small-scale implementations, and gradually increase their use. Provide employees with training and resources, encourage team communication, and incorporate a "human-in-the-loop" workflow where a person evaluates the agent's decisions.
How can businesses measure the success of AI agent management?
Measure the success of AI agent management by tracking several metrics: completion rate, the number of steps an agent takes to complete a task, and time saved. Accuracy is also a key metric, which involves checking if the agent followed instructions and whether a human had to make adjustments.
What are some of the security risks associated with AI agents?
One significant risk is excessive agency, which is when an agent acts too independently. They may delete files, share sensitive information, or break operational processes. You can mitigate this risk by limiting the agent's functionality, permissions, and autonomy, and by using rate limiting and authentication.
How does Cloudflare assist with building and deploying AI agents?
The Cloudflare developer platform offers several tools for building and deploying AI agents, including: Cloudflare Workers AI for globally distributed compute power; Cloudflare containers / sandboxes for secure execution environments; and Cloudflare Agents SDK for scalable and secure deployment. Cloudflare also provides MCP servers and portals that enable secure authentication and a single portal for configuring servers.