TrueForge Puts the Agent Harness, Not the Model, at the Center of Cost Control

TrueFoundry has released TrueForge, an MIT-licensed, self-hostable agent harness that puts tool orchestration, context management, subagents, code sandboxes, human approvals, and execution traces into one vendor-neutral layer. The wager is that teams can reduce the cost of running agents without first accepting a weaker model.

An agent harness is the operating layer around a language model. Rather than simply sending a prompt and returning text, it decides what tools an agent can call, what prior context it receives, when it hands work to subagents, where code runs, which actions need a person’s approval, and how a team can inspect the resulting session. TrueForge offers that layer for local or hosted deployment.

The proposed savings come before model selection

TrueFoundry says TrueForge matched Claude Managed Agents using Opus 4.8 on a 14-task enterprise benchmark while consuming about 30% fewer tokens. It also says routing the harness to GLM-5.2 cut cost by roughly 75% while preserving accuracy. Those are vendor-reported results, not an independent comparison, and the 14-task scope limits what they can establish about a broader production workload.

About 30% fewerToken use versus Claude Managed Agents

TrueFoundry says its harness matched Claude Managed Agents on Opus 4.8 across a 14-task enterprise benchmark while using about 30% fewer tokens.

About 75%Reported cost reduction with GLM-5.2 routing

TrueFoundry says routing to GLM-5.2 reduced cost by about 75% while preserving accuracy in that benchmark context.

The mechanism is not magic compression. A harness can reduce waste by controlling the context supplied to each step, selecting tools and subagents, managing the execution environment, and keeping a trace of what happened. That makes the model only one part of an agent’s cost and behavior; the surrounding software determines how often the model is invoked and how much material it reads.

A crowded layer is becoming more explicit

The launch lands as several companies make the persistent machinery around agents more visible. Anthropic has made computer use, browser tools, Skills API, and Files API generally available on its Claude Platform. Its Files API includes expiration controls, a 500-requests-per-minute limit, and up to 1 TB of storage per organization.

OpenAI has also added collaborative editing to ChatGPT Sites, where teammates can share a project while Codex manages git and CI workflows. Separately, Chroma has introduced Foundation as a research preview for agent memory built from prior sessions. These products address different pieces of the same problem: agents need durable state, action controls, and a way to reuse work rather than starting every task from scratch.

What TrueForge bundles

• Tool orchestration and context management for controlling what an agent can access and carry between steps.
• Subagents and code sandboxes for splitting work and executing it in a managed environment.
• Human approvals and traces for placing review points around actions and recording the session.

Open code does not settle the operating question

The MIT license and self-hosting option give teams more control over where this coordination layer runs. But adoption still asks them to choose how they configure context, tools, approvals, model routing, and observability. The release supplies the components; its benchmark does not show which configuration produces the claimed savings for a particular company’s tools, policies, or task mix.

That unresolved implementation work is the real test for TrueForge. The product’s strongest premise is that organizations need not pick one permanent model for every agent task: they can use a shared control layer to manage costs and execution across model choices. Its next evidence needs to show whether that premise holds beyond a 14-task benchmark, especially when agents run longer sessions, use company-specific tools, and require human checkpoints.