The honest state of agents in 2026
Agents work for narrow, well-instrumented, human-in-the-loop workflows. They fail for open-ended, unsupervised, high-stakes workflows. The middle is where teams lose money.
What works
- Support-ticket triage with a defined action space and a supervisor review queue.
- Data enrichment pipelines with schema validation and rollback.
- Internal-tool orchestration where the agent chooses among five to ten pre-approved actions.
What does not
- "Autonomous" outbound sales agents that email customers without review.
- Multi-step research agents with no ground-truth checks.
- Anything where a wrong action cannot be undone.
The instrumentation floor
Every tool call logged, every decision traceable to the prompt and context that produced it. If you cannot replay a failure, you cannot fix it.