Proactive Agentic Orchestration
POrchestra lets sub-agents request mid-run harness changes, lifting long-horizon task success by about 9%.
POrchestra is an on-the-fly agent orchestration framework that lets sub-agents request harness adjustments at any logical point while continuing local execution. On-Policy Orchestration Distillation attributes those revision requests to earlier orchestration decisions and uses them as process supervision with the final reward. On three long-horizon benchmarks covering deep research, dynamic digital assistance, and software engineering, POrchestra improves task success by 9.1% on average versus the strongest baseline. Combined with OPOD, small-orchestrator success improves 10.0% on average versus existing training baselines, with a better cost-performance trade-off.
- POrchestra lets sub-agents request harness changes mid-run while keeping local execution continuous.
- OPOD turns those revision requests into process-supervision signals for the orchestrator.
- Task success rose 9.1% on average versus the strongest baseline across three long-horizon benchmarks.
- With OPOD, small-orchestrator success rose 10.0% versus existing training baselines.
- Code is public at github.com/XueyangFeng/POrchestra.
From the original article
1 min read at arxiv.orgAgentic orchestration delegates tasks to sub-agents with independent contexts, providing a key path to extending the effective execution horizon of large language model agents.
By Xueyang Feng · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.