Retrieval - Pulling data that already exists within systems you have access to. Internal metrics, database records, uploaded files, analytics dashboards. The defining trait: you are not discovering anything new, you are fetching the current state of something you already own.
Research - Going outside your systems to find new information. Web crawls, external searches, querying third-party data sources. The key distinction from retrieval: the data does not live in your ecosystem yet. Research and retrieval also have different failure modes, so misclassifying one as the other will cause problems during builds.
Reasoning - Any step where an LLM is thinking: summarizing, drafting, scoring, classifying, recommending. Quality depends on two things: the data fed in from upstream steps, and the instructions (prompt) you give the node.
Logic - Conditional branching based on what reasoning found. Placed directly after reasoning nodes to decide a path forward (e.g., "if citation share dropped more than X%, proceed; otherwise, end quietly"). Most agents do not need logic nodes, but they make workflows adaptive rather than blind.
Delivery - Sending the final output somewhere outside the workflow: Slack, email, spreadsheet, CMS, webhook. Without delivery, results stay trapped inside the builder. Design tip: think about delivery first because where the output needs to end up and in what format shapes every decision upstream.
Classifying any step in your workflow:
The design-first process:
Designing before you build is not the slow path. A 22-node agent built on the fly took two weeks and broke constantly. The same workflow, mapped out first, took 30 minutes to build and ran without failure.
Retrieval - Pulling data that already exists within systems you have access to. Internal metrics, database records, uploaded files, analytics dashboards. The defining trait: you are not discovering anything new, you are fetching the current state of something you already own.
Research - Going outside your systems to find new information. Web crawls, external searches, querying third-party data sources. The key distinction from retrieval: the data does not live in your ecosystem yet. Research and retrieval also have different failure modes, so misclassifying one as the other will cause problems during builds.
Reasoning - Any step where an LLM is thinking: summarizing, drafting, scoring, classifying, recommending. Quality depends on two things: the data fed in from upstream steps, and the instructions (prompt) you give the node.
Logic - Conditional branching based on what reasoning found. Placed directly after reasoning nodes to decide a path forward (e.g., "if citation share dropped more than X%, proceed; otherwise, end quietly"). Most agents do not need logic nodes, but they make workflows adaptive rather than blind.
Delivery - Sending the final output somewhere outside the workflow: Slack, email, spreadsheet, CMS, webhook. Without delivery, results stay trapped inside the builder. Design tip: think about delivery first because where the output needs to end up and in what format shapes every decision upstream.
Classifying any step in your workflow:
The design-first process:
Designing before you build is not the slow path. A 22-node agent built on the fly took two weeks and broke constantly. The same workflow, mapped out first, took 30 minutes to build and ran without failure.
