Every AI vendor is selling agents. Every AI framework has an agent abstraction. Every startup deck mentions autonomous agents. Most of what is being called an "AI agent" is either a simple pipeline dressed in marketing language or a genuinely agentic system deployed in a context where a simpler pipeline would work better. Understanding what agents actually are — and when they are the right tool — is one of the most valuable distinctions a technical leader can make right now.
What an AI Agent Actually Is
An AI agent is a system where a language model makes decisions about what actions to take, observes the results of those actions, and decides what to do next — in a loop, until the task is complete. The defining characteristic is that the control flow is determined by the model, not by a fixed program.
In a standard AI pipeline, the sequence of steps is predetermined: receive input → call model → validate output → write to database. In an agentic system, the model decides the sequence: should I search for more information first? Should I call this API? Should I ask for clarification? Should I attempt the task now? The model is the orchestrator.
- Tool use: the model can call external functions (search, APIs, databases, code execution)
- Observation loop: the model receives tool outputs and decides next actions based on them
- Goal persistence: the model keeps the original task in view across multiple steps
- Termination decision: the model decides when the task is complete (or when to escalate)
When the Agentic Pattern Actually Works
Agents work well when three conditions hold: the task genuinely requires dynamic tool selection based on intermediate results, the space of possible action sequences is too large to enumerate in a fixed pipeline, and error recovery is possible (the agent can try a different approach if one fails).
- Research and synthesis tasks: gather information from multiple sources, synthesise, handle missing data dynamically
- Multi-step reasoning over unknown data: the agent discovers what data it needs as it reasons
- Complex problem-solving with tool access: code debugging, data analysis, document drafting with research
- Tasks where the path to completion is genuinely uncertain in advance
When Agents Are the Wrong Choice
Agents are not the right pattern for the majority of AI use cases. The agentic pattern adds latency, cost, unpredictability, and debugging complexity. For most structured business workflows, a deterministic pipeline with an LLM at specific steps is more reliable, faster, cheaper, and easier to maintain.
If you can write down the sequence of steps for the task in advance, you do not need an agent. A pipeline with defined steps, where the model handles the language tasks at each step, will outperform an agent in reliability and be half the engineering complexity.
“If you can write down the steps in advance, you do not need an agent. A pipeline will be more reliable and half the complexity.”
The Reliability Problem with Agents
Agents compound errors. Each action the agent takes is a model decision that can be wrong. A wrong decision at step 3 of a 10-step task propagates through all subsequent steps, often in ways that are hard to detect until the final output is evaluated. The more autonomous the agent, the harder it is to catch and correct intermediate errors.
Production-grade agentic systems require: tight bounds on what tools the agent can call and with what parameters, mandatory human approval checkpoints for high-stakes actions, maximum step limits to prevent infinite loops, comprehensive logging of every action and observation for debugging, and rollback mechanisms for reversible actions.