Open-Source TrueForge Agent Harness Matches Claude Managed Agents Performance at Lower Cost

Developers of the open-source agent harness TrueForge benchmarked their runtime against Claude Managed Agents using the DevRev Enterprise-Bench benchmark. Running the same Opus 4.8 model, TrueForge achieved the exact same 11/14 task solve rate while using 63% fewer tokens and cutting per-run costs by 30%. The benchmark demonstrates that optimizing the agent harness—through context compaction and efficient tool loops—can dramatically lower LLM token consumption without sacrificing task accuracy. Furthermore, maintaining a model-neutral harness allows developers to switch to alternative models like GLM-5.2 to achieve comparable accuracy at up to 75% lower cost. TrueForge reduced average tool calls from 32 down to 19 per task by keeping large outputs out of the LLM context and compacting conversation turns. Released under the MIT license, TrueForge supports OpenAI-compatible endpoints, tool/MCP integration, and persistent sessions, though it currently lacks built-in tracing tools and execution sandboxes.