Agent Workflow
An agent workflow is a structured loop that plans, calls tools or models, observes results, and repeats until a stop condition — a pipeline with memory, contracts, and failure handling rather than a single completion.
Definition
Unlike one chat completion, an agent workflow has explicit steps: retrieve, think, act, observe, stop. Tool schemas, budgets (steps/tokens/cost), and escalation paths are part of the definition. ReAct-style traces are one implementation, not the only one.
Why it matters
Desks care about repeatability: the same inputs should produce the same tool sequence and an audit trail. Workflows turn ad-hoc prompting into something you can budget, evaluate, and shut down when it drifts.
Case
A macro agent that retrieves Codex notes, pulls a chart, then drafts a risk note is a workflow. If any step fails silently and the model still “answers,” you no longer have a workflow — you have decorative automation.
How to read it
Map steps, tools, budgets, and stop rules. Ask what happens on empty retrieval, tool timeout, or contradictory evidence. If those paths are undefined, do not trust the output in production.
Key papers
Landmark references with real publisher or arXiv links — not token-overlap noise.
- ReAct: Synergizing Reasoning and Acting in Language ModelsYao et al. · 2023 · ICLRReason + act loop that most tool agents still copy.