Tencent's WorkForge scales verifiable training environments for long-horizon work agents
teortaxesTex · x · 2026-10-08
A Tencent team published WorkForge, a framework for synthesizing verifiable training environments for work agents that operate on digital artifacts.
- Problem: Hand-crafted agent environments carry prohibitive engineering overhead, while existing synthesis methods sacrifice workspace complexity, realism, or grounded verifiability.
- Approach: Starting from expert workflows, WorkForge retrieves real-world files into a workspace, extracts concrete checkable factual anchors, and derives task instructions, solution plans, and both programmatic and semantic verifiers from them — keeping verification traceable to observable workspace evidence.
- Significance: Offers a scalable path to RL environments for long-horizon professional work agents; the sharer notes Tencent is one of few companies consistently publishing on environment synthesis.
More from coding & agent
- Prime Intellect researcher details experiments on cooperative communicating agent swarms — teortaxesTex · 2026-10-08
- River open recipe: post-trained open models beat GPT-6 Astra Pro and Opus 5.5 on text-to-SQL for under 1% of cost — kylekosic · 2026-10-08
- Agent harness study: prompts and tool schemas drive 3x cost gap on identical models — dair_ai · 2026-10-08
- Anthropic, Google, OpenAI and 7 more offer free official AI courses, incl. full MCP track — anthara_ai · 2026-10-08
- Haiku 5.5 costs 20x less than Sonnet 5.5, dev uses it for parallel sub-agents — RLanceMartin · 2026-10-08
- MCP server drives a real vehicle CAN bus — agents can read and transmit live frames — sa1vatorre · 2026-10-08