Title: From "It Runs" to "Industrial-Grade": Robustness Pitfalls and Engineering Practices in AI Agent Delivery

The most dangerous phase in the delivery process within an AI Lab is the two weeks following a "successful demo."

Many teams, when delivering AI Agents, are accustomed to proving their capabilities through a few carefully designed test cases in an ideal environment. However, once deployed in a real production environment—facing unstructured user inputs, unstable API responses, and complex long-chain dependencies—the originally "smart" Agent often collapses rapidly.

This article shares three pitfalls we encountered and the corresponding engineering solutions we implemented while moving an enterprise-grade knowledge base Agent from prototype to production.

1. Engineering the Truncation of "Hallucinations": From Prompt to Verification Layer

The early approach involved continuously optimizing the System Prompt, instructing the model to "answer 'I don't know' if you are unsure." However, this proved extremely unstable in practice.

Practical Solution: Introduce a [Verification-Correction] Loop.
We no longer trust the model's one-shot output. Instead, we pass the output to a lightweight Verifier.
- Structured Validation: Force the model to output JSON and use Pydantic for schema validation. If validation fails, trigger a retry with error information immediately, rather than exposing the error to the user.
- Citation Traceability Validation: For RAG scenarios, the verifier checks whether the answer generated by the model contains real document fragment IDs. If the answer includes keywords or facts not present in the documents, the response is marked as "low confidence," triggering a secondary retrieval or guiding the user to rephrase their question.

2. Handling "Long-Tail" Exceptions: The Return of the State Machine

The core of an Agent is planning, but pure LLM-based planning is prone to "logical drift" in complex chains. For example, when executing steps A -> B -> C, the model might get stuck in an infinite loop at step B or skip C entirely and declare completion.

Practical Solution: Constrain LLM Freedom with a State Machine.
We define the Agent's behavior as a set of finite state transition graphs (DAGs). The LLM no longer decides "what to do next," but rather "which action to take in the current state."
- Deterministic Paths: Critical business nodes (such as payments and permission checks) are controlled by hard-coded state machines.
- Probabilistic Filling: LLM freedom is granted only in areas like content generation and intent recognition.
This architecture, characterized by a "deterministic skeleton and flexible flesh," reduced the system crash rate by over 60%.

3. Balancing API Instability and Token Costs

In production environments, API timeouts and rate limits are commonplace. Simple try-except blocks cannot resolve cascade failures under high concurrency.

Practical Solution: Multi-Level Degradation Strategy.
- Model Tiering: The primary path uses high-performance models like GPT-4o. When high latency is detected or quota limits are reached, the system automatically switches to lightweight models like GPT-4o-mini or locally deployed Llama-3 to handle simple tasks.
- Semantic Caching Layer: Introduce a vector database to cache answers to common questions (Semantic Cache). If the cosine similarity between a new request and a cached request is > 0.95, the cached result is returned directly without calling the LLM. This not only reduces costs but also lowers the Time to First Token (TTFT) from seconds to milliseconds.

Conclusion

Delivering AI Agents is not a competition of "intelligence," but an engineering practice focused on "certainty." True industrial-grade delivery means: using the most advanced models to provide the upper bound of capability, while employing the strictest engineering measures to secure the lower bound.

Do not attempt to solve stability issues by increasing the length of prompts. Instead, build an engineering system that is fault-tolerant, observable, and rollback-capable to support intelligence.