Workflows vs. Agents

Key idea: A workflow runs fixed steps: cheap and predictable, but it quietly mishandles jobs it wasn't built for. An agent adapts, but costs more and can wander. Most products route routine work to workflows and the rest to an agent.

Jobs done right vs. tokens per job

WorkflowAgentWorkflow + agent fallback
0/105/1010/100307613920Tokens per job (input + output)Run an approach to plot it here
  1. Run Workflow. The routine jobs go fine, but read the replies to the two-part jobs. What did it quietly skip?
  2. Run Agent. Does it handle the two-part jobs? Now check the routine ones and the tokens per job.
  3. Find a routine job the agent got wrong. What did it do instead of the obvious step?
  4. Run Workflow + agent fallback. Where does its dot land between the other two, and why?
  5. Which jobs in your product are routine enough for a workflow, and which need an agent?
AgentoffJob10 support requestsTriagesort + pull out detailsFixed stepscode, per kind of jobReplyto the customerCheckerwas it done?
Press Run to watch it run

Results

Run the jobs to see what each approach did.

Behind the scenes

Run it and every job, the steps taken, and whether it was done will show up here.