𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 𝗛𝗮𝘃𝗲 𝗔 𝗥𝗲𝗹𝗶𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗣𝗿𝗼𝗯𝗹𝗲𝗺
AI agents are moving from software that responds to software that acts. They call APIs, move money, and update databases.
But there is a massive gap between intelligence and reliability.
We focus on better models and better prompting. We ignore the infrastructure. This mismatch causes real-world failures.
Imagine an agent processes a refund. It calls the payment API. The API succeeds. Then, a server crash happens before the agent records the success. The system retries the task. The agent calls the API again. The customer gets a double refund.
No one wrote a bug. The model reasoned correctly. The API worked. The failure happened because the infrastructure is incomplete.
Most agents work fine in demos. Demos run in a single process. They run one task at a time. They do not face crashes or concurrency. Production is different.
When you move agents to production, three things break:
• Process Imortality: Agents assume the process never dies. In reality, hosts die and deployments happen. When a process dies, the in-memory state vanishes. • Pure Tool Calls: Developers treat tool calls like simple reads. But agents perform side effects. Moving money or sending emails cannot be undone easily. • Exactly-once Execution: Retries are necessary for reliability. But retrying an in-memory loop without a durable log creates duplicate actions.
This is not a prompting problem. It is a distributed systems problem. To fix this, we need durable execution.
Reliable agents need these five pillars:
- Event Sourcing: Store an immutable log of every action. The log is the source of truth, not the in-memory state.
- Replayable Execution: Use the log to reconstruct state after a crash. Replay completed steps instead of re-running them.
- Durable Queues: Move work from memory to persistent stores.
- Idempotency Keys: Ensure performing an action twice has the same effect as performing it once. This prevents double payments.
- Compensation Patterns: Define actions to undo steps if a multi-step workflow fails halfway.
A better model produces better decisions. But a better model cannot fix a crash. Reliability is a property of execution, not a property of decisions.
The agents you can trust to act without human oversight will not just be the smartest. They will be the ones running on reliable infrastructure.
智能决定做什么。基础设施确保其能够被正确地执行。
来源：https://dev.to/code_with_mwai/ai-agents-have-a-reliability-problem-nobody-is-talking-about-j40
可选学习社区：https://t.me/GyaanSetuAi