AgentSkillsCN

production-operations

通过“观察—行动—进化”循环，动态管理运行中的智能体。借此保持智能体的性能稳定，管控不可预测的成本，并基于生产数据，有针对性地优化智能体的表现。

SKILL.md
--- frontmatter
name: production-operations
description: managing live agents through the Observe-Act-Evolve loop. Use this to maintain performance, manage unpredictable costs, and strategically improve agents based on production data.

Production Operations for Agents

Goal

Establish a continuous operational model that manages the inherent autonomy of agents, keeping them reliable, cost-effective, and safe as they interact with real-world users.

The Continuous Operational Loop

Unlike static services, agents require an integrated cycle of intervention:

1. Observe (The Sensory System)

Gain deep insight into the agent's internal "thought process" using three pillars:

  • •Logs: Granular, factual records of every tool call, error, and decision.
  • •Traces: Narrative threads that reveal the causal path of why an agent took a certain action.
  • •Metrics: Aggregated reports on performance (latency), cost (token count), and operational health.

2. Act (Tactical Reflexes)

Real-time levers to stabilize the system:

  • •Scaling: Decouple logic from state. Use stateless, containerized services with externalized state management (e.g., Vertex AI Agent Engine's session service) to scale horizontally.
  • •Reliability: Implement automatic retries with exponential backoff for failed calls. Ensure tools are idempotent to prevent duplicate actions (like double-charging) during retries.
  • •Security Containment: Use "circuit breakers" (feature flags) to instantly disable tools if a threat is detected.

3. Evolve (Strategic Improvement)

Proactively fix root causes identified in the "Observe" phase:

  • •Data-Driven Refinement: Analyze production failures to create new, permanent test cases for your evaluation dataset.
  • •Rapid Deployment: Use an automated CI/CD pipeline to commit refined prompts, new tools, or updated guardrails and deploy them in hours or days rather than months.

Optimization Levers

  • •Speed: Work in parallel and use smaller, efficient models for routine tasks.
  • •Cost: Shorten prompts, use cheaper models for easy steps, and batch requests where possible.
  • •Granularity vs. Overhead: Set a lower default log level (INFO) in production and use dynamic sampling (e.g., trace 10% of successes but 100% of errors).