Summary
Sam Witteveen introduces TrueForge, an open-source MIT-licensed managed agent execution harness from TrueFoundry that lets developers run production-grade agents entirely on their own infrastructure — without being locked to a specific model provider or paying hosted runtime premiums charged by Claude Managed Agents, OpenAI’s agent API, or Gemini Managed Agents.
The video opens with a useful taxonomy of the three existing agent harness categories: coding harnesses (Claude Code, Codex, OpenCode), assembly-yourself frameworks (LangGraph, Agent SDK, ADK), and hosted runtime harnesses (the managed agent tier). TrueForge represents a fourth option — an open-source equivalent of that third category — built in TypeScript, deployable via a single `npx` command, and backed by a local SQLite database that keeps all session data on-premise.
Witteveen walks through TrueForge’s core capabilities: session persistence, streaming, MCP and skill integration, sandboxed tool execution, human-in-the-loop approvals, a test-bench chat UI, and an HTTP API plus TypeScript SDK for embedding agents into Slack, mobile apps, or any external software. The tool is particularly relevant for enterprise and regulated-industry teams that face compliance mandates preventing data from leaving their own cloud — and for developers running open-weights models locally via Ollama or LM Studio. The video explains why harness choice matters at volume: the execution harness, not just model token cost, often determines how many tool calls an agent makes and how efficiently it reaches a result.
📺 Source: Sam Witteveen · Published September 14, 2026
🏷️ Format: Tutorial Demo
