TrueFoundry's TrueForge Open-Sources an Agent Runtime That Cuts Claude Costs by 75%

Building an agent that demos well is easy. Keeping one alive in production is where most teams stall, and the layer where they stall has a name: the harness. TrueForge, just open-sourced by TrueFoundry under an MIT license, is a full runtime for that layer, and it lands with benchmarks claiming it can match Claude Managed Agents on the same model while cutting costs by 30-75%.

The runtime layer nobody wants to write twice

An agent harness is the loop that sits between the model and the outside world. It runs the full execution loop, planning, tool routing and execution, context management for long tasks, security boundaries like sandboxing and human-in-the-loop approvals, and session state that survives reconnects and restarts. Writing this yourself means reinventing streaming, tool servers, approval gates, and audit trails before you even get to the interesting part of your agent.

TrueForge ships as three parts: a core server that runs the agent loop, an HTTP API (with a TypeScript SDK) to drive it from code, and a chat UI (with a React UI SDK) to drive it from the browser. Everything available in the UI is exposed through the API, so any workflow you would otherwise click through can be automated.

Sandbox as a tool, not a prison

The most interesting design choice sits in how code execution is handled. Many harnesses run the agent inside a sandbox. TrueForge treats the sandbox as a tool: one gets provisioned only when the agent actually needs to execute code. That is why one server can run many agents concurrently, and why turns that skip code execution are cheaper and faster.

Sandboxing today runs on Daytona, with more providers planned. Skills, which are git-backed SKILL.md instruction packs loaded on demand inside the sandbox, only get mounted when needed. The theme is aggressive laziness: nothing gets loaded, provisioned, or paid for until the agent actually uses it.