TrueForge: Open Source AI Agent Harness Cuts Costs 30-75%

What happened

TrueFoundry, a San Francisco-based B2B machine learning startup founded in 2021 by former Meta engineers, has released an open source AI agent harness called TrueForge under the permissive MIT License on GitHub. The tool is designed to give developers and enterprises greater control over how AI agents select and use tools, while significantly reducing the cost of running them.

In a blog post, TrueFoundry reported that when TrueForge was paired with the open source GLM-5.2 large language model, it successfully completed 11 of 14 tasks on DevRev's Enterprise-Bench — a benchmark that tests multi-step tool use across CRM, issue tracking and document management systems. According to the company, this result was achieved at 75% lower cost than reaching comparable outcomes with Anthropic's Claude Managed Agents harness running on Claude Opus 4.8. TrueFoundry states that, more broadly, TrueForge delivers task completion at 30% to 75% lower cost than the Claude-managed alternative.

Because TrueForge is released under the MIT License, it can be freely forked, modified, self-hosted and paired with any preferred model, rather than being tied to a single vendor's managed infrastructure.

Why it matters

This is fundamentally a technology story about the maturing economics of AI agent infrastructure. As enterprises move from pilot projects to production deployments of AI agents, the "harness" — the orchestration layer that governs how an agent plans, calls tools and manages context — is becoming as consequential a decision as the underlying model itself. TrueForge's open, model-agnostic design signals a shift away from single-vendor managed agent stacks toward architectures where organisations can mix and match models and infrastructure to control cost.

For technology and transformation leaders, the emergence of credible open source alternatives to proprietary managed-agent offerings widens the field of viable options for scaling agentic AI. It suggests that cost-per-task, not just raw model capability, is becoming a key competitive battleground — one that will shape how organisations budget for and architect AI-driven automation across customer-facing and back-office workflows alike.

By the numbers

• 30-75% lower cost claimed for TrueForge task completion versus Claude Managed Agents, according to TrueFoundry
• 75% cost reduction reported in the specific DevRev Enterprise-Bench comparison against Claude Opus 4.8
• 11 of 14 tasks successfully completed by TrueForge paired with GLM-5.2 on DevRev's Enterprise-Bench
• 2021 founding year of TrueFoundry, by former Meta engineers

The Renascence take

The headline number here is cost, but the more interesting signal is control. Enterprises adopting agentic AI are discovering that vendor lock-in at the orchestration layer can be as limiting as it once was at the cloud or CRM layer — and an open, self-hostable harness is a direct response to that anxiety.

Most coverage of agent benchmarks fixates on the percentage saved, but the real story is architectural leverage: an open harness lets an enterprise swap models as pricing and capability shift, rather than being captive to one provider's roadmap and rates. The behavioral lesson for buyers is that agentic AI decisions should be made in layers — model, harness, governance — not as a single bundled purchase. Operators piloting AI agents for service or operations should benchmark harnesses against their own task mix before committing, since a 75% saving on one benchmark's tasks won't automatically translate to every workflow.