AI agent workflows
Applied AI workflows that match each task to a specialized role and model.
A large task can contain planning, execution, and review work. I use a specialized workflow so those jobs have different responsibilities instead of becoming one vague instruction.
The public account stays at the pattern level. It explains how I think about the work without exposing private implementation details or making a performance claim I cannot support.
- TIER
- 1 / EVIDENCE-LIGHT
- FORMAT
- TEXT-FIRST
- SURFACE
- TEXT-FIRST CASE STUDY
Problem
Planning a task, doing the task, and checking the result are different kinds of work. Treating them as one step makes it harder to see what needs attention.
Observation
Specialized roles help when each role has a clear job and the handoff between them is explicit. The workflow is only as useful as that separation.
Personal action
I use a pattern with specialized planning, execution, and review roles, then match the task to the tool or model that fits it.
Constraints
This page stays at the approved pattern level. It leaves private infrastructure, costs, prompts, secrets, tool names, model names, and unsupported performance claims out.
Result
The workflow makes responsibility easier to read: planning frames the task, execution does the work, and review checks what came back.
Learning
Specialization helps only when the handoffs are clear. More roles do not fix a workflow that cannot say who owns the next decision.
Change
I now match the job to the role before I choose how to run it. That keeps the method flexible without making the responsibility fuzzy.