Three practical agent archetypes
| Pattern | What it does well | Common failure mode |
|---|---|---|
| Planning agent | Breaks goals into sequenced sub-tasks and checkpoints | Over-plans without execution deadlines |
| Retrieval agent | Finds and grounds responses in enterprise sources | Poor source ranking creates confident but wrong synthesis |
| Action agent | Performs operations through approved tools/APIs | Unsafe side effects if permissions are too broad |
Guardrails you should deploy from day one
- Scoped permissions with task-level allowlists
- Human approval for irreversible actions
- Structured output schemas and policy validation
- Comprehensive run logs for incident review
Ship strategy: Start with assistive mode, prove accuracy and containment, then graduate to supervised autonomy by workflow tier.