Outcome. You can trace an agent run, identify its state and stopping rule, and locate the step where an error, loop, or false completion should have been caught.
A chatbot usually maps one request to one response. An agentic system lets a model participate in a loop: observe the current state, choose an action, receive the result, update state, and continue until a stop condition is met. The model may plan or select tools, but ordinary software still manages messages, permissions, timeouts, retries, memory, and the final handoff.
Agency is a spectrum. A fixed workflow with model-powered steps is often easier to test than an open-ended agent that chooses its own path. More autonomy can help when the route is genuinely uncertain, but it also expands the space of possible failures. A good system has step and cost budgets, explicit success criteria, recoverable state, structured errors, and escalation when confidence or permission is insufficient.
“Done” must be verified against external state. A model saying “the file was uploaded” is not proof; the application should check the upload result or storage record. Otherwise a fluent completion message can hide a failed action.
Mental model. The model proposes the next move; the harness owns state, limits, execution, verification, and stopping.
Evidence trail — reviewed 23 July 2026. ReAct formalizes interleaved reasoning, actions, and observations: https://arxiv.org/abs/2210.03629. Anthropic’s engineering guidance recommends starting with the simplest measured workflow: https://www.anthropic.com/engineering/building-effective-agents.
Add state containing queries already tried, sources accepted/rejected, remaining evidence gaps, step count, and cost. Stop successfully only when three sources pass criteria; stop unsuccessfully after six steps or repeated no-progress; ask the user when the goal is ambiguous; verify that every cited URL was actually retrieved.