Building Responsible AI Agents in Decentralized Open-Source Environments
summary
A minimal on‑prem pipeline is built and demoed end to end using LLaMA, Gemma, GPT‑OSS plus MCP, A2A, AutoGen and Google ADK with vLLM, Hugging Face Transformers and Presidio for privacy tooling A custom framework called Agent of Sealed runs agents as agentized services across two machines (16 GB and 6 GB GPUs) using model offload/onload routing and a policy-as-code guard that votes across models Password generation, SQL injection threat analysis and a calendar agent are implemented and mostly succeed with real API rate limits and occasional network hiccups.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 14:20Topic overview
- 17:12Cloud-first to data-first AI
- 21:31Modular policy-driven agent stack
- 29:36Decentralized compute and open-source models
- 34:11Agent framework demo
- 53:57Guardrails and safety mechanisms
- 1:05:41Use case walkthrough
- 1:20:10Calendar agent demo
- 1:35:35Q&A
related talks
30mMind the Gap - A Developer's Roadmap To Building AI Agents
30mBuilding Reliable AI Agents: An Open Source Approach To Evaluation and Observability
40mAgents in Action: Advancing Open Source AI Missions
37mAGENTS OF S.E.A.L.E.D: AI Agentic Cybersecurity Framework
21mDemocratizing AI with Open Source Multi-Agent Systems
01h 27mBuilding Vertical Agent with AG2 AgentOS
24mFrom First Experiments to AI Agents: Adoption Experience and Future Challenges
32mProgress In the Decentralized AI Stack