Agent Infra智能体基础设施

A 12-skill checklist collected from the original notes.从原始笔记整理的 12 项技能清单。

Runtime & orchestration运行与编排

Agent Runtime智能体运行时
Understand agent loops, tool calling, streaming and sessions. Build a minimal agent that can work through a task in a loop.理解 Agent Loop、Tool Calling、Streaming 和 Session；实现一个能够循环执行任务的最小 Agent。
Context & state management上下文与状态管理
Distinguish the context a model can see from a task’s execution state. Handle context trimming, compression, checkpoints, short-term context and long-term memory.区分模型当前能看到的上下文与任务执行状态。处理上下文裁剪、压缩、Checkpoint，以及短期上下文和长期记忆。
Task queues & workers任务队列与 Worker
Distribute tasks through queues. Manage concurrency, priorities, backpressure and worker pools.通过 Queue 分发任务，控制并发、优先级、Backpressure 和 Worker Pool。
Scheduling & multi-agent orchestration调度与多 Agent 编排
Break complex tasks into subtasks. Coordinate DAGs, reviewers, handoffs and human-in-the-loop workflows.拆分复杂任务，组织 DAG、Reviewer、Handoff 和 Human-in-the-loop。

Durable execution持久化执行
Persist execution state. Use checkpoints, replay, resume and recovery to continue long-running tasks after a process restart.持久化执行状态，通过 Checkpoint、Replay、Resume 和 Recovery，在进程重启后恢复长任务。
Retries, timeouts & idempotency重试、超时与幂等
Set retry and timeout policies for model, tool, network and worker failures. Prevent side effects such as code pushes, emails and resource creation from running twice.为模型、工具、网络和 Worker 的失败设置重试与超时，避免推送代码、发邮件、创建资源等操作重复执行。
Distributed state & consistency分布式状态与一致性
Understand transactions, WAL, leases, heartbeats, replication, consistency and recovery. Coordinate task state across orchestrators, workers and databases.理解 Transaction、WAL、Lease、Heartbeat、Replication、Consistency 与故障恢复，协调 Orchestrator、Worker 和数据库的任务状态。

Sandboxes & execution environments沙箱与执行环境
Isolate each agent’s filesystem, processes, network and resources. Limit CPU, memory and execution time.为 Agent 隔离文件系统、进程、网络与资源，限制 CPU、内存和执行时间。
Tool infrastructure & MCP工具基础设施与 MCP
Manage shell, filesystem, browser, database and API tools through a common layer. Understand MCP as an interface for tools and resources.统一管理 Shell、Filesystem、Browser、Database 和 API 等工具；理解 MCP 作为工具与资源接口的作用。
Permissions, security & secrets权限、安全与密钥
Control access to files, tools, networks and secrets. Understand authentication, authorization, ACLs, secret management and auditing.控制文件、工具、网络与密钥访问，理解 Authentication、Authorization、ACL、Secret Management 和 Audit。

Observability & debugging可观测性与调试
Trace the full execution of a task. Use metrics to monitor latency, queues, tokens, cost, failure rates and retries.用 Trace 记录完整执行过程，用 Metrics 观察延迟、队列、Token、成本、失败率和重试。
Evaluation & replay评估与回放
Evaluate results continuously with datasets, rubrics, task success rates, regression checks, replay and LLM-as-judge methods.通过 Dataset、Rubric、Task Success Rate、Regression、Replay 和 LLM-as-Judge 持续评估任务结果。