What an agent actually is (and isn't)

"Agent" is the most overloaded word in AI. Strip the hype: an agent is software where a language model decides, in a loop, what to do next.

The minimum viable definition

An agent is a system with:

1. A goal (often specified per-invocation).
2. A set of actions it can take (tools, APIs, code execution, human messaging).
3. A loop where the LLM picks the next action, observes the result, and decides again.

Everything else — planning, memory, multi-agent coordination — is optimization of that core loop.

Why the word gets abused

"Agentic" gets slapped on any software that uses an LLM. A chatbot isn't an agent. A single-turn summarizer isn't an agent. A workflow with one LLM call in the middle isn't an agent. A loop makes it an agent.

The three failure modes of the loop

Every agent failure looks like one of these:

1. Wrong next action. The model chose a tool that doesn't get closer to the goal. Root cause: ambiguous tool descriptions or missing context.
2. Infinite loop. The model keeps trying the same thing. Root cause: poor error handling — the model doesn't learn from failed tool calls.
3. Premature termination. The model declares success when the goal isn't met. Root cause: no verification step.

Most of agent engineering is mitigating these three.

What agents are good at (honestly, in 2026)

• Multi-step information gathering. "Research X and produce a brief." Strong.
• Software engineering on bounded tasks. Open a PR to fix a well-specified bug. Getting strong.
• Customer support triage and partial resolution. Strong for common flows, weak at edge cases.
• Web navigation for well-defined transactions. "Book this for me" is still error-prone; "find me three options and summarize" is solid.

What agents are still bad at

• Tasks requiring stable long-term memory across many sessions.
• Anything with high stakes per action (financial transactions, medical orders, irreversible data operations) without heavy human oversight.
• Novel tasks with no precedent in training. The loop doesn't magically invent new methods.

The designer's job

When you build an agent, your job is:

• Make the goal specific enough to verify.
• Make the tools few and well-described.
• Make the loop safe — able to recover from failures, able to ask for help, able to stop.

Give a model a vague goal and a bag of tools and you get a demo. Give it a precise goal and three reliable tools and you get something that ships.