Agent
A model given tools, a goal and permission to keep going on its own.
A system in which an LLM decides what to do next, calls tools to act, observes the results and repeats until the task is done or it is stopped. The model provides judgment; the surrounding harness provides everything else.
"Agent" is a spectrum, not a binary. Autocomplete is not an agent; an assistant that proposes one change you approve is a little bit of one; a process that reads a ticket, edits files, runs tests and opens a pull request while you sleep very much is.
The definition worth defending is the dynamic one: an agent decides its own next step, where a workflow follows a path you wrote. That distinction matters because it predicts the failure modes. Agentic systems fail in ways you cannot enumerate in advance, which is why autonomy and verification have to be designed together.
Choosing a workflow when a workflow will do
A document pipeline always does the same three things in the same order.
Write it as a chain, not an agent. The dynamic loop buys flexibility you do not need, and pays for it in latency, cost and unpredictability. The simplest sufficient design is the one you can debug.
Autonomy is a dial, not a category
Deciding how much rope to give an agent on a production repository.
Start read-only on the real system while it drafts, then allow edits behind review, then allow unattended work only in a sandbox with automated checks. Each step up is justified by what the previous level demonstrated.
Calling a fixed prompt-and-response chain an "autonomous agent". If there is no loop and no tool choice, it is a workflow.
- blogBuilding effective agents ↗
Draws the line between fixed workflows and agents that choose their own process, and names the core patterns: chaining, routing, parallelisation, orchestrator-workers, evaluator-optimizer.
- reportAgents ↗
A conceptual white paper on agent components and architecture — model, tools, orchestration layer — widely cited as the vendor-neutral starting frame.
- paperReAct: synergizing reasoning and acting in language models ↗
Interleaves reasoning traces with environment actions so observations can correct the plan. The paper every agent loop is a descendant of.
Agent Loop · Agentic loop
Think, act, observe — repeated until the task is finished.
Harness · Scaffold
Everything around the model that makes it an agent.
Tool Call
The model asking for a specific function with specific arguments.
Human in the Loop · HITL
A person reviews or approves before the work lands.
AFK · Away from keyboard
Running without a person watching.