XNews

TrueForge Harness Anatomy: How Long-Running Agents Survive Beyond Demos

TrueForge Harness Anatomy: How Long-Running Agents Survive Beyond Demos

Long-running agents depend on the harness, not just the model. TrueForge, an open-source harness, illustrates the architecture: model, MCP tools, skills, sandbox, plus sub-agents, context management, approval checkpoints, and persistent sessions. The execution loop requires handling infinite tool calls, large results, and browser crashes. Progressive disclosure loads tool schemas on demand. Skills separate “how” from “what.” Sandboxes isolate execution and offload context. Context management uses progressive disclosure, code-mode filtering, externalizing large results, and compression. Sub-agents enable parallel research with isolated contexts. Approval gates are harness-level hard rules, not prompt requests. Event streams persist every step, allowing disconnect-resume and full replayability.

Source posts