Sandbox demo

What Makes a System Agentic

The loop that separates agents from workflows — watch one run, then learn when NOT to build one.

A workflow is code you wrote calling an LLM at fixed points — the control flow is yours. An agent flips it: the model directs its own process, choosing tools and actions in a loop against a goal, reading each result to decide the next step. The loop below is the whole trick: model → tool call → result → model, until it decides it's done. Everything fancy — multi-agent systems, computer use, deep research — is this loop with better tools and guardrails.

The honest engineering rule: use the least agency that solves the problem. Fixed steps? Write a workflow — it's cheaper, faster, and debuggable. Bounded branching? A router plus workflows. Reserve true agents for open-ended tasks where you genuinely can't enumerate the path: debugging, research, multi-step operations over unpredictable state. Agency costs tokens, latency, and failure modes that compound per step; spend it where it buys capability.

Sandbox transcript
Claude Code · sandbox replay
Press Play to replay this transcript step-by-step.
0 / 8

The transcript is canned — no real API calls. The point is to internalize the loop: user → plan → tool → result → answer.

Check your understanding
Q1. In the replay above, what made the system agentic rather than a workflow?
Q2. A task has three fixed steps that never vary. You should build:
· Score 100% on the quiz.