Image by Author showing 7 roles/personalities of AI Agents You would not send a single junior employee to build an entire software product from scratch. No requirements gathering, no architecture review, no QA, no nothing. Just one person, alone, figuring it all out. That would be a disaster. Yet that is exactly how most people are building AI agents today. One model. One prompt. Hope for the best. The agents that actually work in production look nothing like that. They look like teams. The Solo Agent Illusion There is a common misconception that a powerful enough LLM, given the right prompt, can handle complex end-to-end tasks on its own. And for simple tasks, that is sometimes true. But complexity breaks the illusion fast. Think about building a mobile app. There is user requirements gathering, architecture planning, code generation, testing, and publishing. Each of these steps demands different skills, different focus, and different judgment. No single model does all of that well in one shot. LLMs are increasingly being used in agentic settings in which they must take sequences of actions or must plan and make a series of decisions to complete longer-horizon tasks. — Anthropic, “Claude’s Character and Core Views,” 2024 Longer-horizon tasks are not a prompt engineering problem. They are an architecture problem. And the solution, it turns out, looks a lot like how human organizations have always solved complexity: by building teams. Think of It Like a Startup A scrappy startup does not hire one person and expect them to do sales, engineering, design, and finance simultaneously. You hire the right people for the right roles. Same logic applies to multi-agent systems. When you design an agentic AI system, the first two questions are: what roles do you need, and how do you make each role genuinely good at its job? Here are the core roles that show up again and again in well-designed agent teams. The Six Roles Every Agent Team Needs The Doer This is your execution layer. In a human team, think junior developer or data entry specialist. The doer handles individual, well-scoped tasks: writing a function, formatting a document, making an API call. On its own, a doer cannot build an app. But with someone else holding the bigger picture, it can execute steps reliably. The Planner Before any code gets written, someone has to break the problem down. The planner takes the user’s input and turns it into a structured sequence of steps. In the mobile app example, that means one planning pass for user requirements and a second for app architecture, before a single line of code is generated. The planner does not execute. It thinks. Chain-of-thought prompting enables LLMs to decompose complex problems into intermediate reasoning steps, significantly improving performance on multi-step tasks. — Wei et al., “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models,” NeurIPS 2022 The planner is the chain-of-thought in team form. The Tool Operator Most real-world tasks require reaching outside the model: calling APIs, querying databases, running code, interacting with web services. The tool operator handles this. It structures the inputs, fires the tool call, and feeds the results back into the workflow. This is the bridge between the language model and the actual world. The Learner Your agent often needs information it was not trained on. A learner gathers that. It might retrieve competitor product data, scan recent blog posts for user sentiment, or pull from internal documents using retrieval-augmented generation (RAG). The skill here is not just fetching data. It is knowing what is relevant and what is noise. Retrieval-Augmented Generation (RAG) combines the parametric memory of a pre-trained model with a non-parametric memory, allowing the model to access current, specific, or private information dynamically. — Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks,” NeurIPS 2020 The Critic Every team needs the person who points out what is wrong before it ships. In an agent team, the critic reviews outputs for hallucinations, runs QA tests on generated code, or scores multiple candidate outputs and selects the best one. Healthy competition between agents, just like healthy competition in a team, produces better results. Self-refinement allows models to iteratively improve their own outputs through a feedback loop, without requiring additional training data or human supervision. — Madaan et al., “Self-Refine: Iterative Refinement with Self-Feedback,” NeurIPS 2023 The Supervisor As workflows grow more complex, someone needs to track whether things are actually moving forward. The supervisor monitors individual roles for stalls, checks progress at the task level, and flags failures before they cascade. It is the agent equivalent of a project manager who notices when a team member is stuck on day two instead of day ten. The Presenter At the end of any project, someone has to synthesize and communicate. The presenter pulls together the work of the whole team and communicates it back to the user in a clear, useful format. In the mobile app example, that means summarizing requirements, explaining what the codebase does, and telling the user what was actually built. The ReAct Pattern: A Simple Starting Point If you have heard of the ReAct pattern, you have already seen a compact version of this team. ReAct combines: A Reason step (Planner) An Act step (Tool Operator) An Observe step (Critic) An Answer step (Presenter) ReAct prompts LLMs to generate both verbal reasoning traces and actions in an interleaved manner, allowing the model to reason about what to do, act on that reasoning, and observe the result. — Yao et al., “ReAct: Synergizing Reasoning and Acting in Language Models,” ICLR 2023 ReAct is a great starting point. But as your tasks get more complex and your reliability requirements go up, you need more specialization, not just more prompting. Making Each Role Actually Good at Its Job Designing the roles is only half the work. The other half is making each role genuinely effective. There are four levers. Prompting is the most accessible. Just as you give clear instructions to a new hire, you give clear, role-specific instructions through the system prompt. Simple things matter: “If you get stuck, retry.” “Focus only on this step.” Model Selection is about fit. Not every role needs the largest, most expensive model. A critic role might benefit from a reasoning model. A doer handling simple transformations might work fine with a smaller, faster one. Match the model to the task, not just the budget. Model Tuning is the most resource-intensive option. Fine-tuning requires labeled examples of both good and bad outputs, plus compute to adjust model weights. It pays off when a role needs to be highly specialized or highly consistent. It is not where most teams should start. Context Management is often underestimated. Giving an agent access to the right information at the right time, and keeping irrelevant information out, is one of the highest-leverage design decisions you can make. Overloaded context degrades performance. Focused context sharpens it. The performance of LLM agents is highly sensitive to what information is included in the context window. Irrelevant or excessively long context can hurt performance, while well-curated context consistently improves task success rates. — Liu et al., “Lost in the Middle: How Language Models Use Long Contexts,” TACL 2024 Start Small. Grow Deliberately. You do not need to build all six roles on day one. A two or three role agent, a planner, a doer, and a presenter, can ship real value fast. Think of it like the early days of a startup: a small, focused team moves quickly. But just as a startup eventually fills gaps, adds specialists, and builds internal processes to maintain quality at scale, your agent team needs to grow too. The goal is not just “does it work once.” It is “does it work consistently, across a wide range of inputs, under real-world conditions.” That is when feedback loops, supervision, and specialization start to matter. What This Means for You If you are building AI agents today, stop thinking in prompts and start thinking in teams. Map your workflow. Identify the distinct roles that each step actually requires. Choose the right model for each role. Build feedback mechanisms in from the start. Complex tasks need more than a smart model. They need a smart system. Faisal Feroz is a Chief Technical Architect and Fractional CTO with 20+ years of experience building enterprise AI platforms, event-driven architectures, and GenAI systems at scale. He regularly writes on the practical realities of AI architecture, agentic systems, and engineering leadership. Follow his work on LinkedIn and read more on his blog . The Future of Agentic AI is Not One Genius Model, it is a Team was originally published in Artificial Intelligence in Plain English on Medium, where people are continuing the conversation by highlighting and responding to this story.