The Agent Is a Workflow, Not a Prompt

Learn AI from practitioners by studying the workflow, reconstructing the smallest useful experiment, and measuring results before copying the tool stack.

A practitioner maps an AI workflow with visible checks, traces, and human review.

The fastest way to learn AI is not to collect more tools. It is to study how strong practitioners turn a vague task into a bounded workflow, add the right tools and checks, and inspect the failures. This series will follow that practice across blogs, X posts, videos, talks, repositories, and demos.

The signal

The method in one sentence

Start with one useful workflow, test its parts and its full trajectory separately, then improve from traces and failure cases.

What the sources actually show

The production pattern has four layers:

  1. Define a task and a success condition that a human can inspect.
  2. Test individual components—prompt, retrieval, tool call, parser, or guardrail.
  3. Test the complete multi-step trajectory, including wrong turns and recovery.
  4. Move from sandbox to a small canary only after the evidence is good enough.

This is a stronger learning loop than asking an agent to “do the whole thing” and judging one impressive answer.

Reconstruction plan

  • Task: turn one research question into a source-backed brief.
  • Baseline: a single prompt to a general model.
  • Smallest useful experiment: compare the baseline with a two-step workflow: researcher gathers sources and claims; reviewer checks citations, missing alternatives, and unsupported conclusions.
  • Artifact to save: the prompt, source list, draft, review notes, and final brief.
  • Success condition: every important claim has a source, uncertainty is visible, and the reviewer catches at least one weakness the baseline missed.
  • Stop condition: stop adding agents if the second step does not improve the artifact or if review cost exceeds the value of the task.

Reality check

  • Observed: the sources recommend component-level evaluation, trajectory analysis, and staged rollout.
  • Reconstructed: not yet run in this first field note.
  • Reported: the production guidance comes from the authors' own engineering experience.
  • Still unknown: which workflow, model, and review threshold will work best for our research writing.
  • Constraint: more agent steps create more latency, failure surfaces, and maintenance work.

Transfer rule

When learning from a practitioner, copy the decision structure before copying the framework or model: task boundary, tools, checks, artifact, and failure response. If those are unclear, the source is inspiration—not yet a method.

Practice card

  • Inputs: one research question, five candidate sources, one baseline prompt.
  • Steps: baseline answer → source-and-claim extraction → reviewer pass → compare artifacts.
  • Expected output: a brief with claims, links, confidence, and unresolved questions.
  • Quality check: inspect citation coverage and the changes made by the reviewer.
  • Next iteration: add one failure case from the first run to the evaluation set.

For the next field note, see Find the First Wrong Step.

Chinese companion: 智能体:工作流,而非提示词.