Human in the Loop

A Human You'll Need in the Loop: The Agentic Engineer

• By Howard M. Cohen
• January 20, 2026

Just over a year ago, we were talking about citizen developers, people with knowledge of their own business processes using low-code/no-code (LCNC) tiled interfaces to assemble software needed to support those processes by simply moving tiles around the screen. We often talked about how this took a substantial load off software developers' plates, freeing them for higher-level development tasks.

Much has changed since then. Citizen developers are now, by and large, completing their transition from the LCNC platforms to using popular AI chatbots like ChatGPT, Gemini, Claude, and others to create much of what they need, from simple business functions like writing emails, producing reports, creating marketing materials, simple workflows, and similar tasks.

Popular AI Misconceptions

Many people think the chatbot is the entirety of how people use AI. They know they can ask questions and get answers. Some refer to AI as "search on steroids," mainly because that's all they're aware of.

Much has been written and said about the fear many people have of AI: that, as it becomes more and more autonomous, there's a danger it will take over the world and wipe out the human race. You may find this hard to take seriously, but among those warning about this is Geoffrey Hinton, widely considered "The Godfather of AI."

The AI Continuum

Perhaps the most generic misconception is that AI is a single thing existing at a single level. Actually, AI can be said to exist at several levels across a broad continuum. These levels include:

• Generative – Users instruct their AI chatbot to create various creative content, including images, videos, music, models, and more.
• Assistive – The AI, usually a chatbot, assists the user in performing various tasks, performing research, summarizing large documents, and writing various documents.
• Collaborative or Copilot – The AI, still starting with a chatbot, uses tools to create various kinds of content in collaboration with the user. One popular example is using Anthropic Claude to create new Microsoft Excel spreadsheets. Claude is so adept at this that Microsoft decided to change the large language model (LLM) underlying Copilot for Microsoft 365 from OpenAI's ChatGPT to Claude.
• Autonomous – Using a dedicated interface or an integrated development environment (IDE), the user instructs an AI Agent to perform various tasks that may require multiple steps, including examination, evaluation, decision-making, planning, and taking action to make necessary changes in the real world. Basic autonomous AI agents often interact repeatedly with the human in the loop for additional information and decision support.
• Fully Autonomous – The role of the Human-in-the-Loop is solely to review results and approve or make change orders. The fully autonomous AI agent continues cycling, awaiting specific trigger events to resume its operation.

The Autonomy Paradox

When people like Godfather Hinton talk about their fear of AI takeover and human demise, they are referring to these fully autonomous agents. The fear is that the lack of human supervision will leave these agents free to commit all manner of horrors.

The paradox is that, according to industry studies, less than 1% of all current deployments are fully autonomous AI agents. In other words, the thing we're most afraid of is something nobody can find good use cases for.

The AI Skills Gap

As they scale from chatbots to developing Agents, they find they have people who can "chatting" with AI, but what they really need are engineers who can architect and build the AI-based systems their company needs. Citizen Developers needed Software

Developers for more complex projects, and they still need someone in that role. Thus, the AI question in most companies is now shifting from "How do we get our people to use AI?" to "Who is architecting the agents that will run our operations?"

These Agentic AI Systems bring several requirements, including:

• Perception & Grounding (The "Eyes and Ears")
  A true agentic system doesn't just wait for text input; it actively perceives its environment. Using Multimodal Input, it can ingest not just text, but images such as screenshots of errors, files in PDF, CSVs, MD, and other formats, and system logs.The AI understands where it is. It knows, "I am operating in the Production environment," or "I am looking at the Finance SharePoint." It does not hallucinate that it is in a void; rather, it is "grounded" in your specific business reality. This is referred to as Environment Grounding.
• Planning & Reasoning
  The single biggest differentiator from a standard chatbot is that the system must be able to "think before it acts."

Task Decomposition: The capability to take a vague goal and break it down into a step-by-step plan.

Chain of Thought (CoT): The system must generate a hidden "internal monologue" where it justifies its decisions before executing them.

Self-Correction: If a step fails, the system must have the logic to retry or try a different method, rather than just crashing or immediately asking the user for help.

• Tool Use & Action (The "Hands") An agentic system must be able to change the state of the real world, not just describe it. The system connects to the Model Context Protocol (MCP) to read and write data. It doesn't just say "I sent the email"; it actually calls the email API to send it. It also has a "menu" of tools such as a Calculator, Calendar, CRM, and Email and the intelligence to pick the right tool for the right moment without human intervention.o assure safety, it includes "human-in-the-loop" checkpoints for high-stakes actions such as approving refunds or agreements.
• Memory & State Management: Standard LLMs generally have amnesia. That is, they forget everything once the chat window closes. Agentic systems must remember.
• In the short term, it remembers its position in a multi-step workflow. For long-term persistence in use, retrieval augmented generation (RAG) stores knowledge in a vector database so it can recall facts from months ago. It also logs its own past successes and failures to learn what worked previously.
• Critique & Reflection
  Reliable agents have a built-in quality control layer. Before showing an answer to the user, a secondary internal process reviews it to assure it fully answered the user's inquiry and presented the response in the appropriate format. It also has hard-coded rules that override the model if it tries to do something that is forbidden, such as deleting data or sharing private information.

The "Agentic Loop"

A system is only "agentic" if it can autonomously loop through this cycle:
Perceive → Plan → Act (Tool Use) → Reflect (Did it work?) → Remember.

The Agentic Engineer

This need for Agentic AI Systems creates a necessary new role with a new skill set, the Agentic Engineer.

You should transition from "Citizen Development" to assigning an "Agentic Engineer" when your needs meet any of these three criteria:

• Complexity: The task requires more than 3 steps of logic (e.g., "Read email, look up sender in Salesforce, check inventory, then reply"). Citizen developers struggle to make this reliable.
• Connectivity: The solution requires writing data back to a system of record. Giving a chatbot "write access" without an engineer's security architecture (MCP) is a massive risk.
• Consistency: You need the solution to work 99% of the time, not 80%. "Prompt engineering" alone hits a ceiling; "System engineering" is required to break through it.

Requirements for the Agentic Engineer

This is a specialized technical role responsible for architecting and assembling AI systems. Unlike a "normal user" who chats with a chatbot, the Agentic Engineer:

• Orchestrates Context: They meticulously engineer what data is fed into the model to ensure accuracy, reduce hallucinations, and moderate operating expense.
• Builds "Scaffolding": They build the code and logic around the LLM. This includes connecting RAG pipelines, defining tools via MCP, and managing multi-step workflows where one AI agent passes tasks to another.
• Moves from "Chat" to "System": They transition from asking questions in chat to designing repeatable, reliable systems through engineering.
• Advanced Context Engineering: Design and optimize extensive system prompts and context injection strategies to ensure high-fidelity model responses.
• RAG Pipeline Architecture: Build and maintain RAG systems to ensure the AI can accurately retrieve and synthesize proprietary data.
• Agentic Tooling & MCP: Implement MCP to securely connect LLMs to internal APIs, databases, and third-party software, enabling the AI to read and write data rather than just passively retrieving information.
• Evaluation & reliability: Create automated test suites, called "evals", to measure the accuracy and reliability of AI outputs before deployment thus moving from more casual "vibe coding" to engineered reliability.

The role of the Agentic Engineer is still being actively defined, so the terminology is undetermined. One common term applied to this role is AI Orchestrator. This emphasizes the skill of managing multiple models, APIs, and data sources to work in harmony.

The Agentic Engineer is responsible for designing, building, and refining complex AI workflows that go beyond simple chatbots. This role bridges the gap between business needs and raw AI capabilities by engineering the context, connectivity, and control required for enterprise-grade solutions. The Agentic Engineer you engage will need the following skills:

• Fluency in "Mid-Code": Proficiency in Python or TypeScript to write the "glue code" that connects LLMs to the real-world using frameworks like LangChain, AutoGen, or raw API calls.
• Orchestration Logic: Ability to design multi-agent flows where a "Planner Agent" delegates tasks to "Worker Agents".
• Security & Governance: Understanding of prompt injection risks and data privacy boundaries when connecting AI to business data.

The New Essential Role

The Agentic Engineer is the bridge between the raw potential of Frontier Models like ChatGPT, Gemini, or Claude and the specific reality of your business data. They are the builders who take perception, planning, action, and memory and weave them into software that doesn't just chat, but actually does work.