Agent runtime

Agentic Workflow

A workflow where one or more AI agents make decisions about what to do next — choosing tools, gathering context, escalating to humans — rather than executing a fixed script.

Operating principle

Production AI is not a prompt. It is a system of context, tools, permissions, traces, evals, and feedback loops.

What it is

An agentic workflow is a process where part of the control flow is decided by an AI model at runtime rather than encoded ahead of time. Classic workflows have a fixed graph: 'step 1, then step 2, then step 3'. Agentic workflows have a graph plus a model that decides which step to take next based on what it just saw. The result is more flexible — and harder to make reliable.

Why it matters

Many real-world processes don't fit a fixed script: support triage, research, debugging, complex form-filling. Agentic workflows handle those by letting the model adapt. The trade-off is determinism: a fixed workflow always does the same thing; an agentic one might not. Production deployments depend on traces, evals, and approval gates to make agentic behavior accountable.

How it works

The agent loop: receive event, assess context, decide on next action (call a tool, ask a clarifying question, escalate, finalize), execute, observe, repeat. Common patterns: ReAct (reason then act), plan-and-execute, supervisor with sub-agents. Substrates: LangGraph, OpenAI Agents SDK, Mastra.

Related resources