From Copy-Paste to Composition: Building Agents Like Real Software
summary
Claims that agent runtimes need an adaptation layer not another giant prompt and demonstrates a working virtual tools layer that intercepts MCP tool lists via an MCP proxy, a central registry, and lightweight workflows (multi‑source search calling Exa/GitHub/HuggingFace, Haiku LLM, a graph DB triple store). Shows deterministic wiring cuts LLM calls dramatically, reports an 80% success horizon trend and a practical demo at scale inside an enterprise that ultimately completes the composed workflow.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:22The evolving AI landscape
- 3:14Natural language programming
- 4:39The agent-tool coupling problem
- 7:39The virtual tool solution
- 10:19Workflow composition
- 13:17Architecture integration
- 14:29Demo: research assistant
- 21:15Governance and adoption
- 23:00Q&A and closing
related talks
18mExposing Agents as MCP servers with mcp-agent
40mBuilding Agents with Agent Bricks and MCP
31mThe Future Is Domain-Specific Agents
40mHow To Tame Your Agents?
01h 03mAI Coding Agents and how to code them
25mBeyond the IDE: Toward Multi-Agent Orchestration
32mProgramming language evolution in the AI era
16mBeyond the IDE: The Future of Developing Software With Agents