Workflow Fundamentals¶
A workflow is a graph of steps. Some steps are code, some are an LLM call, some are a full agent loop. The first design decision is always: how much of this graph should the model control?
flowchart LR
J["Junior: trace one step"] --> M["Middle: pick a pattern"]
M --> S["Senior: draw the determinism boundary"]
S --> P["Professional: govern the autonomy budget"]
Levels¶
| Level | Guide | You are done when |
|---|---|---|
| Junior | Trace one step, then a chain | You can tell a single LLM call apart from an agent loop, and write down a step's input/output contract. |
| Middle | Pick the workflow pattern | You can choose among the 5 canonical patterns and justify it against latency, cost, and predictability. |
| Senior | Draw the determinism boundary | You can say exactly which parts of a workflow the model decides and which parts code decides, and why. |
| Professional | Govern the autonomy budget | You can set org-wide rules for how much autonomy a new workflow is allowed to start with. |
Practice rule¶
Before adding a branch, a loop, or a sub-agent to a workflow, write down: "the model decides X here because code cannot." If you can't finish that sentence, that piece belongs in code, not in the model's hands.
Related¶
- Orchestration and Delegation — once a workflow needs more than one agent, this is how they hand off work.
- Reliability and Recovery — what happens when a step in this workflow fails or a human needs to approve it.
- Tools and MCP — what a step actually calls when it acts.