ProgAgent: training agents with programmatic tool calling beats direct tool calling
hkust-nlp · hf · 2026-10-15
HKUST researchers introduce ProgAgent, a full training stack for Programmatic Tool Calling (PTC), where tools are invoked via executable code and intermediate results live in runtime variables instead of the context. The stack includes a pipeline for synthesizing verifiable long-horizon tasks, a lightweight PTC runtime, and a two-stage training recipe. Using only 215 instantiated tasks, ProgAgent beats comparable-size baselines across four benchmarks; on LOCA-bench, ProgAgent-14B hits a 30.7% pass rate at a 64k-token environment state where baselines score zero. Cross-harness experiments show PTC-trained models outperform DTC-trained ones under both evaluation interfaces, suggesting the training benefit goes beyond the interface.
More from coding & agent
- One-man software factory: 5-10 parallel agents build POCs, 60% get killed — Stefania_druga · 2026-10-15
- Huawei's SwarmRSI: evolving agent swarms achieve recursive self-improvement in science — Qianru Meng · 2026-10-15
- Prompt With Love: Steering Massive Agent Swarms With Values, Culture and Story — liminal_bardo · 2026-10-15
- Vibe First, Validate Fast: One Chart Nails How AI Development Differs From Agile — burny_tech · 2026-10-15
- A Year of Agent Memory Evolution: From Static CLAUDE.md to "Dreaming" — mcraddock · 2026-10-15
- patio11 on multi-agent dev: clearly the future, and we'll buy so much inference — IgorCarron · 2026-10-15