TrueForge Open Source Agent Harness Cuts Costs 75%

TrueFoundry open-sourced TrueForge on August 19, 2026, an MIT-licensed agent harness built around a specific thesis: most of what an AI agent costs is not the reasoning, it is the context you keep resending. The benchmark numbers the company published make that case in the bluntest possible terms.

• On DevRev's Enterprise-Bench, TrueForge with the open-weights GLM-5.2 model completed tasks for $2.90 versus $11.80, a 75% reduction
• Holding the model constant at Opus 4.8 on both sides, TrueForge still came in 30% cheaper at $8.50 versus $11.80
• Released under the MIT license, free to fork, modify, and use commercially
• Deployable locally with SQLite, or in production via Docker Compose and Helm with Postgres and Redis

What Is an Agent Harness?

A harness is the loop around the model. It decides which tools the agent can see, how tool results come back, what stays in the conversation, and when to stop. The model does the reasoning; the harness does the bookkeeping. Two harnesses running the identical model can differ enormously in cost and reliability, which is exactly what the Opus 4.8 comparison isolates.

TrueFoundry was founded in 2021 by former Meta engineers Nikunj Bajaj, Abhishek Choudhary, and Anuraag Gutgutia, and the company sells enterprise machine learning infrastructure. Open-sourcing the harness while keeping the platform commercial is a familiar strategy, and in this case the harness is genuinely usable standalone.

Where Does the 75% Saving Come From?

Four techniques, all of them variations on sending the model less. TrueForge delays loading tool schemas until they are actually needed, rather than paying for every tool definition on every turn. It delegates work to subagents so a long subtask does not bloat the parent conversation. It offloads large tool results to files and passes references instead of contents. And it compacts the conversation automatically once it crosses a 50,000-token threshold.

None of these are novel individually. What is notable is the measured effect of applying them systematically, and that TrueFoundry published a like-for-like comparison rather than only the favourable model-swap number. The 30% figure at identical model choice is the honest measure of the harness itself; the 75% figure combines harness efficiency with the cheaper open-weights model.

Why Context Engineering Is the Current Frontier

The past year of agentic AI development has quietly shifted from "which model reasons best" to "how little can we send it." Context windows kept growing, and teams discovered that filling them is both expensive and, past a point, actively harmful to accuracy. Harnesses are where that discipline gets encoded.

It is a space with real competition and real experimentation, including approaches where the agent rewrites its own scaffolding, as we covered with Xiaomi's HarnessX self-improving harness. TrueForge takes the opposite tack: deterministic, auditable context management that a platform team can reason about.

What Ships in the Box

Beyond the context work, TrueForge runs the full agent execution loop — model calls, MCP tools, skills, sandboxing, approvals, and session state. The release includes over 40 built-in tools, sandboxed execution, human-in-the-loop approvals, generative UI streaming, and Tavily-powered web search, distributed on both GitHub and PyPI.

Vendor neutrality is the other design commitment. Any model can be plugged in, and models can be swapped per task to trade latency against quality against cost — which is precisely how the GLM-5.2 result was achieved. TrueFoundry names NetApp and Automatiq as early users, and is launching a hosted pay-per-usage version alongside the open-source release. For enterprises weighing agent deployments, an MIT-licensed harness that runs on your own Postgres is a straightforward thing to evaluate, and open-weights models like Ornith-1.5 scoring 86.1 on Terminal-Bench give it capable engines to drive.

Sources: TrueFoundry — August 19, 2026; VentureBeat — August 19, 2026; Open Source For You — August 2026.