@Jerry_fsy: If you're interested in AI-related work, you might consider the emerging field of Agent Infra. I've recently been getti…

If you're interested in AI-related work, you might consider the emerging field of Agent Infra. I've recently been getting started in this area myself, conducted some research, and compiled a list of skills required in this field—feel free to bookmark and save it. I'll also continue sharing my learning insights in this domain. Full List: 1. Agent Runtime: Understand Agent Loop, Tool Calling, Streaming, Session; understand and implement a minimal viable Agent, rather than just calling the model API once. 2. Context and State Management: The context fed to the large model determines what the model can see in this round, and it largely determines the quality of the model's decisions; State will save where the task is currently executing. Need to handle context trimming, compression, Checkpoint, task state, long-term memory, and short-term context. 3. Task Queue and Worker: Distribute tasks to Workers for execution via Queue, and control concurrency, priority, Backpressure, and Worker Pool. 4. Scheduling and Multi-Agent Orchestration: Break down complex tasks into multiple subtasks, supporting DAG, Reviewer, Handoff, and Human-in-the-loop. 5. Durable Execution: Even if the Agent runs for tens of minutes or even hours, it won't lose the task due to process restarts; requires persisting execution state, Checkpoint, Replay, Resume, and Recovery. 6. Retry, Timeout, and Idempotency: Models, tools, networks, and Workers will all encounter issues; the system needs to determine when to retry, when to timeout, and avoid duplicate executions of operations like git push, sending emails, or creating resources. 7. Distributed State and Consistency: Multiple Orchestrators, Workers, and databases jointly maintain Task State; need to understand Transaction, WAL, Lease, Heartbeat, Replication, Consistency, and fault recovery. 8. Sandbox and Execution Environment: Coding Agents will execute Shell, Git, Python, builds, and tests; need to provide isolated file systems, processes, networks, and resource environments for each Agent, and limit CPU, memory, and execution time. 9. Tool Infrastructure and MCP: Agents aren't just calling a few Functions; they also need unified management of tools like Shell, Filesystem, Browser, Database, API; MCP can serve as a standard interface between Agents and external tools/resources. 10. Permissions, Security, and Secrets: Control which files Agents can access, which tools they can call, which networks they can connect to, and which keys they can read, involving Authentication, Authorization, ACL, Secret Management, and Audit. 11. Observability and Debugging: The full execution process of Agent Tasks will be recorded via traces, using Metrics to monitor latency, Queue, Token, cost, failure rate, and Retry. 12. Evaluation and Replay: Not just checking "if the Agent can run," but also determining if the task was truly completed correctly, through Dataset, Rubric, Task Success Rate, Regression, Replay, and LLM-as-Judge for continuous evaluation.