Why Human‑Designed Task Planning Beats Fully Autonomous AI Agents in Enterprise Databases

Introduction

AI agents are built on a simple architecture: Agent = LLM + Plan + Memory + Tools . Phenomenal agents such as DeepResearch, Manus, and Claude Code all follow this framework.

Task Planning Is Critical

The quality of the planning step directly determines the effectiveness of the final answer; a good plan can make a small model outperform a larger one.

Should Planning Be Fully Autonomous?

There is an ongoing debate: should large models handle the entire planning process, or should humans define the plan while the AI only executes and analyzes? This question impacts both technical complexity and business outcome.

Industry Forecast

According to Gartner’s 2024 report, by 2028 33 % of enterprise software applications will embed Agentic AI (less than 1 % in 2024), and at least 15 % of daily work decisions will be made autonomously by agents .

Case Study: Alibaba Cloud RDS AI Assistant

We opened source code for Alibaba Cloud RDS MCP in April 2025. The system prompt emphasizes “Task decomposition first: must provide detailed steps”. Our experiments showed that fully autonomous agents often hallucinate and achieve only ~20 % accuracy, whereas manually‑planned agents reach >85 % accuracy on the same set of issues.

Why Enterprises Prefer Stability Over Cleverness

Enterprise deployments care about reliability, not raw intelligence. Reliability includes:

Explainability: the agent must show reasoning and data sources.

Repeatability: the same scenario should yield the same result.

Accuracy: the agent must avoid hallucinations and provide trustworthy conclusions.

In database operations, these requirements translate into a strict SOP workflow: collect metrics → locate root cause → generate remediation suggestions . For example, when CPU load spikes, the agent automatically checks connections, slow queries, and index usage before proposing fixes.

Prompt Example

# Role: Database SQL Issue Diagnosis Expert
You are a professional database diagnostic specialist. Your goal is to break down the problem, call the appropriate tools, and provide accurate, data‑driven recommendations. Do not fabricate information.

## Workflow
1. Get time‑range information.
2. Extract slow‑query logs.
3. Retrieve table schema if needed.
4. Check error logs.
5. Gather performance monitoring data.
6. Analyze monitoring metrics.
7. Summarize findings and give optimization advice.

**First, decompose the task and briefly describe each step, then follow the workflow.**

Hybrid Planning with a Case Library

Instead of maintaining an exploding rule base, we built a case library derived from ten years of RDS operational SOPs and thousands of tickets. When a new issue arrives, the system matches it against similar cases, then follows the matched workflow. This approach keeps the system flexible while preserving explainability.

Multi‑Agent Architecture

We designed three agent types:

Exploratory Agent: large‑model‑driven planning for open‑ended queries, with safety prompts to limit hallucinations.

Execution Agent: strict SOP‑driven planning for high‑frequency, deterministic tasks (CPU spikes, storage exhaustion, slow SQL).

Hybrid Agent: switches between autonomous and manual planning based on keyword routing and intent classification.

Keyword‑based routing (e.g., "CPU使用率", "慢SQL", "磁盘空间") directs the request to the appropriate agent, dramatically reducing context length and improving tool‑call accuracy.

Practical Tips

Inject the current timestamp into prompts to avoid relative‑time misunderstandings, and explicitly list available tools to prevent the model from inventing nonexistent ones.

Conclusion

AI agent planning should not be viewed as an either‑or choice. For enterprise‑grade scenarios that demand stability, explainability, and high accuracy, human‑crafted planning—augmented by large‑model execution—remains essential. The RDS AI Assistant demonstrates that combining domain expertise with LLM capabilities transforms agents from “chat toys” into reliable, production‑ready tools.