Towards Useful, Reliable, and Scalable AI Agents
summary
IBM Research releases an open-source Bee stack for building production agents: B agent framework (TypeScript first, Python coming), B code interpreter (sandboxed Python on Kubernetes), and an observability pipeline that logs events into a store and exports to mlflow. Demonstrated an agent that calls two tools (web search and Python) to compute US population in 2030 and returned a numeric result, and showed runtime observability and JSON/Zod-driven schemas. Key focuses are llm drivers per model, guardrails for prompt injection and data leaks, serialization, and determinism tradeoffs.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:03Defining LLMs and chat LLMs
- 5:40Pitfalls of large language models
- 8:16What is an LLM agent
- 11:07LLM agent challenges
- 15:24Safety, security, privacy
- 18:40B open-source stack
- 21:56B agent framework deep dive
- 25:09Building agent tools
- 26:57Standalone LLMs
- 28:04Structured outputs with drivers
- 29:50Observability in framework
- 31:37Serialization and state
- 32:43Roadmap and features
- 33:26Q&A
related talks
39mTowards Useful, Reliable, and Scalable AI Agents
28mBuilding and Sharing Reliable AI Agents with BeeAI
30mBuilding Reliable AI Agents: An Open Source Approach To Evaluation and Observability
40mAgents in Action: Advancing Open Source AI Missions
25mDirecting a Swarm of Agents for Fun and Profit
41mThree agents, three frameworks, one talk
45mAI agents in practice - beyond the prototype
40mAI Agents in Practice