'Agents' are the current hype cycle, and behind the noise there are real, paying use cases and real traps. The difference is whether the workflow is bounded, tool-using, and tolerant of occasional error caught by review. Here's the honest map.
Where agents pay off
- Multi-step but bounded tasks: triage, enrichment, routing, drafting.
- Tool use: querying systems, filling forms, calling APIs under supervision.
- High volume, low individual stakes, where a review step catches misfires.
- Internal workflows where you control the environment and data.
Where they fail
- High-stakes actions with no review, irreversible or costly errors.
- Ambiguous goals or sprawling context the agent can't reliably hold.
- Workflows needing guarantees, where 'usually right' isn't good enough.
- Anything where a deterministic tool would simply be better.
