ZenML 新项目 Keturu 聚焦持久 AI Agent 构建
从 MLOps 到 Agent 工程，核心原则仍是可靠性；ZenML 推出 Keturu 实践这一理念。
主持Daniel Whitenack
嘉宾Hamza Tahir
话题脉络
- 2:48MLOps principles carry over
MLOps learnings from productionizing ML applications translate directly to building durable AI agents.
Hamza - 5:49Agents as unrolled graphs
An agent is just a loop of LLM and tool calls, essentially an unrolled workflow graph.
Hamza - 8:46Why agents lack durability
Agents often lack durability because they run locally without scaling infrastructure to handle failures or high throughput.
Hamza - 12:50Battle for the agent stack
The industry is split between application-level and infrastructure-level agent builders; the winner is unclear.
Hamza - 16:49Harness beats model quality
A well-designed agent harness (tools, meta-prompts) can outperform a better model, enabling an RL feedback loop.
Hamza - 22:32Practical fragility of agents
Agent deployments face nontrivial challenges like persistence, failure management, and long-running tasks.
Daniel - 27:09Building durable agent platforms
Investing in state management, worker resilience, and versioning is essential for scaling enterprise agents.
Hamza - 35:22Traceability for optimization
New SDK provides traceability and re-evaluation of agent runs, enabling data-driven optimization.
Hamza - 41:25Open models encourage open harnesses
Open model progress fosters internal platforms and open harness standards, reducing model lock-in.
Hamza
节目简介
ZenML 联合创始人 Hamza Tahir 从 MLOps 视角切入，认为构建可靠 AI Agent 的核心并非全新框架，而是复用软件工程中的确定性、可重试原则。其新项目 Keturu 专门针对 Agent 的持久化问题，让非确定性代码在生产中安全运行。