TrueFoundry, a San Francisco ML startup founded by former Meta engineers, released TrueForge under the MIT License this week. On DevRev's Enterprise-Bench, 14 multi-step tasks spanning CRM, issue tracking, and document management, TrueForge paired with the open source GLM-5.2 LLM completed 11 tasks at $2.90 total. The same benchmark using Anthropic's Claude Managed Agents with Claude Opus 4.8 cost $11.80, a 75% premium. Using identical models in both harnesses, TrueForge still runs 30% cheaper: $8.50 versus $11.80.
The cost gap comes from context engineering, not model swaps. TrueForge delays loading MCP tool schemas until needed, offloads oversized tool results to files instead of the active context window, delegates isolated tasks to subagents, and compacts conversations automatically at a default threshold of 50,000 tokens. The sandbox is provisioned only when the agent needs to execute code, not kept running throughout the entire task loop. TrueFoundry's COO Anuraag Gutgutia is explicit about the business logic: TrueForge is not a replacement for Claude Managed Agents or other harnesses. It is the open layer beneath them, designed to route all enterprise agent traffic through TrueFoundry's paid AI Gateway, which handles model access, credentials, permissions, budgets, and observability.
The full article is worth reading for the harness comparison table, which maps TrueForge against DeepSeek's harness, OpenAI Codex CLI, LangChain Deep Agents, and Claude Managed Agents across license, price, model flexibility, and deployment options. Claude Managed Agents charges $0.08 per running session-hour on top of token costs. TrueForge charges nothing for the harness itself. Deployment scales from a single local process with SQLite to Docker Compose or Helm with Postgres and Redis, though TrueFoundry explicitly flags the local config as unsuitable for internet-facing production use.
[READ ORIGINAL →]