Give Your Agent a Computer
summary
Claim: giving an agent a file system changes behavior — it follows through on long tasks and builds on prior work. Uses AI SDK v6 with a tool-loop agent, provider-executed web search, Vercel persistent named sandboxes, a bash tool, a memories.md file and generated Python scripts. Built system runs long sessions (one demo 104 minute run, 316 tool calls, 29 files changed) and drops support tickets ~90% while getting ~95% positive thanks. file system as memory bash tool + memories.md
AI-written summary · about
chapters (18)
- 0:00Setup and project initialization
- 7:31Installing dependencies
- 9:02Introduction to AI SDK v6 and tool loop agents
- 11:05Defining the agent in agent.ts
- 13:40Building the route handler for the chat API
- 15:26Creating the frontend UI with useChat
- 16:53Modifying agent instructions
- 18:40Adding web search tool (provider-executed tools)
- 23:55Adding UI components for tool usage
- 27:43The importance of persistent sandboxes (file systems for agents)
- 33:45Managing context and message history across steps
- 45:11Initializing persistent sandboxes
- 47:20Defining custom call options and context
- 51:10Creating the bash execution tool
- 55:05Integrating the bash tool into the agent
- 57:15Adding persistent memory (memories.md)
- 1:03:44Making agents learn by generating Python scripts
- 1:05:58Reviewing the full agent system and sub-agents
related talks
33mAgent Experience
44mHow Does a Coding Agent Work?
39mAgents for the Sake of Happiness
58mAdvanced Autonomous AI Agents: Memory, Collaboration and Evolution in Action
30mBeyond the Hype: Using AI Coding Agents Today
24mAI Agent Swarms Are Amazing
49mBuilding a coding agent from scratch
49mLet's build an AI agent