Tencent Hunyuan's WorkForge builds 16.7K verifiable real-world environments for work agents
TencentHunyuan · x · 2026-10-22
Tencent Hunyuan introduced WorkForge, a framework that constructs verifiable training environments for long-horizon work agents from real-world resources.
Key points:
- Starting from expert workflows, it builds authentic workspaces (profiles with summaries, facts, schemas), extracts checkable facts, and derives tasks with dual verifiers: programmatic checks plus semantic judges grounded in observable evidence
- Scales to 16.7K environments across 40 professional domains and 60 file types (tabular files dominate at 53.31%, images 18.69%)
- Real-world workspaces are far harder than synthetic ones: 13× larger (279KB vs 21.5KB), 5.4× more files per workspace, 2.4× file-type coverage
Gains on long-horizon work benchmarks grow consistently with data volume and interaction horizon; Bash (149K calls), Edit, and Read are the most-invoked tools.
Related event: Tencent Hunyuan Open-Sources WorkForge for Long-Horizon Agents(2 posts)→
More from coding & agent
- Red Team Pass prompt: make the model critique and fix its own output — thisguyknowsai · 2026-10-22
- Assumption Audit prompt: list assumptions first, then confirm before answering — thisguyknowsai · 2026-10-22
- Source Challenger prompt reportedly cuts fabricated citations by 67% — thisguyknowsai · 2026-10-22
- Confidence Scorer prompt: make models label every claim HIGH/MEDIUM/LOW — thisguyknowsai · 2026-10-22
- Refusal Protocol prompt: make models admit UNCERTAIN to cut hallucinations 41% — thisguyknowsai · 2026-10-22
- Boundary Setter prompt: hard constraints over soft instructions — thisguyknowsai · 2026-10-22