Anti-Patterns
Common ways agent projects become vague, unsafe, or hard to recover in production.
Most failed agent implementations are not model failures. They are design failures.
Avoid these patterns
The general-purpose do-everything agent
If one agent is expected to answer questions, take actions, coordinate systems, and invent its own workflow, it will be hard to reason about and harder to trust.
Hidden side effects
If a step can perform writes or external actions without that being obvious from the design, replay and debugging become risky.
Vague success criteria
“Handle customer requests” is not a workable contract. Define what good output looks like and what should happen when the agent is unsure.
Unbounded tool access
Broad tool permissions create accidental power. Keep the allowed surface small and aligned to the job.
No human fallback
If there is no escalation path, the first ambiguous or risky case becomes an operational problem.