It makes sense then that when we talk about ‘AI Agents’ we mirror the range of uses for the word agent in human jobs;
AI Agents: A Practical Framework for Enterprise Automation
Understanding the potential of AI agents - from simple workflows to autonomous collaborators - a framework for talking about and evolving AI agent standards.
Prefer to watch?
Fraser Paine walks through the same framework in more detail:
Agent Levels 1–3
The Three T’s (Tasks • Tools • Triggers)—with concrete examples, adoption pitfalls, and the “don’t buy AI, hire it” playbook.
In ~20 minutes, get the key models, how to scope a pilot, and the steps to take agents from experiment to production ROI.
Introduction
A lot has been said about agents, but definitions vary wildly depending the context and incentives of the person talking. Just about every tech company today wants to claim that their latest product release supports agentic AI, but this makes discussing on clear and agreed terms challenging.
Businesses are starting to ask themselves, ‘If I want to build an agent; should I have a junior engineer whip up an automation workflow in a few days, or have a whole team build robust supporting infrastructure and an extensive code base with monitoring and test coverage to automate an entire role in my organisation?’
This range reflects the reality of work today, some tasks are simple, but most jobs require many tasks, and some jobs require completely open ended skillsets from day to day.
- 1
A Customer Services Agent or CSA answers routine questions or takes small actions on behalf of a customer calling in for support. Not unlike todays AI Chatbots.
- 2
A Travel Agent or Real Estate Agent do a set of tasks that make something complex very simple for the average person, such as booking travel or buying a house. Just like modern AI powered workflows.
- 3
- A Special Agent or a Field Agent in the CIA or FBI might perform a very dynamic set of responsibilities using a wide array of tools to solve problems. Much more like the autonomous agents we are seeing emerge right now.
To help with this confusion we propose a simple framework for describing agents that covers this whole spectrum.
Anatomy of an Agent
Agents are made up of many moving parts, some parts are hard coded logic, others are dynamic context windows or complex prompts. However, all of these components broadly fall into three categories:
- Tasks: The sequence of actions performed by the agent, and the conditions used to choose between different sequences of actions based on the context.
- Tools: The actions available to the agent which either send, generate, or receive data and interface with other systems or agents.
- Triggers: The events which cause the agent to execute a given task in each context.
Each of these components can have varying levels of complexity and maturity in their support of the agents capability, some agents automate extremely complex tasks but are only ever triggered by simple chatbot requests from users, such as OpenAIs Deep Research, others complete very simple tasks but are triggered by using an LLM to monitor unstructured data feeds or messages, emails, and new documents being uploaded.
Here we break down the maturity level for each of these components to help you better understand the complexity of agents you’re thinking of building, or are already working with today.
| Tasks | Tools | Triggers | |
| Level 1 | Tasks are specific workflows or sequences defined by engineers with very few branching conditions | Tools are a defined set of interfaces that execute parameterised queries or actions. | Triggers are timer based or responses to user requests in a chatbot or API |
| Level 2 | Tasks are defined by users but make heavy use of agent determined branching | Tools are defined and documented by users but their parameters and usage are determined by the agent at runtime | Triggers are conditional based on business systems events such as emails with certain context or documents being uploaded to a specific folder |
| Level 3 | Tasks are generated or selected by the agent based on the trigger conditions, context, and goals | Tools are dynamically discovered and written by the agent using API documentation, computer interfacing, and coding tools | Triggers are dynamic and determined by the agent as it monitors a set of live feeds for activities related to its role and functions |
Agent Maturity Case Study: Sales Team
Here I will use the example of a sales team that uses increasingly powerful agents to support their function within their business.
Level 1: Retrieval
A basic workflow that triggers when a user asks a question to a chatbot is a Level 1 retrieval agent. This might support a sales team in retrieving case studies, or information about potential clients, not simply running a search but reasoning over the request and the search results, running follow up searches or calculations, and then returning a concise and accurate answer.
Level 2: Tasks
An agent that is triggered by a user or an event such as an email, which then does a simple task or set of tasks would be a Level 2 agent. This might support a sales team by classifying emails into categories and executing workflows based on the category of email, such as updating a CRM, responding to the sender, and updating an account manager on the activities.
Level 3: Autonomy
An agent that monitors a range of email inboxes, chat streams and online data sources to autonomously understand the environment and then determine a course of actions to support a business is a Level 3 agent. This might support the same sales team by monitoring a public tender service to look for opportunities, then compare these to historical projects in the organisation, use this context to evaluate the opportunity, create a response draft, and then contact team members to review and contribute to the bidding process. This kind of agent needs access to a range of public and private data sources, authoring tools, and an identity with which to integrate with the organisation.
Wrapping Up: Thinking Clearly About Agents
As AI agents evolve from simple workflow automations to dynamic systems capable of complex, goal-oriented action, the need for clarity in how we define and design them becomes critical. Not every agent needs to be autonomous—or even intelligent—in order to deliver business value. But understanding where your agent sits on the maturity spectrum of Tasks, Tools, and Triggers helps you scope the right solution, set appropriate expectations, and plan for what comes next.
Whether you're building a Level 1 assistant to streamline information retrieval, or laying the groundwork for a Level 3 autonomous collaborator, this framework offers a shared language for teams to evaluate opportunities, manage complexity, and make strategic decisions around agent development.
In a world increasingly shaped by agentic AI, knowing what kind of agent you're building, and why, is the first step toward building something that works.
Start Building the Right AI Agent for Your Business
Whether you’re exploring simple workflow automations or designing advanced, autonomous agents, clarity is key. Let’s discuss your goals, assess where your needs sit on the Agent Maturity Spectrum, and map out the right approach to deliver real business value.
Discover more about AI