Open-source TrueForge harness matches Claude managed agents' accuracy with 63% fewer tokens
Background-Job-862 · reddit · 2026-09-03
TrueFoundry open-sourced TrueForge, a model-neutral general agent harness (MIT, OpenAI-compatible endpoints, MCP, subagents, approvals, sandbox integration), and benchmarked it on 14 DevRev Enterprise-Bench tasks with a blind judge:
- Claude Managed Agents + Opus 4.8: 11/14 solved, $11.8/run, 10M tokens/run
- TrueForge + Opus 4.8: 11/14 solved, $8.6/run, 3.7M tokens/run, 19 vs 32 tool calls
- TrueForge + GLM-5.2: 11.7/14 solved, $3.0/run — roughly 75% cheaper
The gains come from the agent loop itself: less carried context, compaction, fewer tool calls, keeping large outputs out of context. Caveats: no first-class tracing/eval, bring-your-own sandbox, lossy compaction. Benchmark methodology is published for reproduction.
More from coding & agent
- LukeW built an agent-orchestration app for a foldable: managing agents is now an all-day job — LukeW · 2026-10-21
- Agentic SDLC: engineers shift from implementers to agent orchestrators as cycles shrink from weeks to hours — Pavan_Belagatti · 2026-10-21
- OpenAI's dots can now use company tools via Merge for Workforce with employee-bound permissions — shensi · 2026-10-21
- The topology of an LLM agent: where the model, loop, and tools run shapes cost and security — davidcrawshaw · 2026-10-21
- Agents with full-disk and accessibility access are an enterprise security nightmare, argues Runlayer — leveragedupside · 2026-10-21
- Acellera runs an unsupervised virtual lab of AI scientists on a real IL-23 molecular discovery program — gdefabritiis · 2026-10-21