Awesome AI Agents › AI Assistants

renee-jia/scholar-loop

⭐ 469 Python repository created 2026-06-15

ScholarLoop automates the working cycle of a research scientist with eight cooperating agents and a deterministic harness around them. A Director reads the run ledger and literature trends to set the next direction, topic and budget. A Lit Scout pulls real papers from arXiv and OpenAlex, ranks them by citation impact and distills cited techniques. A Reasoner combines those priors with past lessons and constraints to propose several distinct experiments, and a three-persona debate panel votes on whether each deserves compute. Surviving ideas climb a funnel of smoke, verify and full tiers, run in parallel up to a worker limit, with weak candidates killed after one cheap run and stronger ones promoted to three and then five seeds with significance testing. A Runner executes genuine PyTorch experiments, currently digit classification scored by error rate and diabetes regression scored by root mean squared error, with new domains added as a configuration profile and an engine pair rather than orchestrator changes. A Reflector converts each outcome into a lesson in a relevance-ranked, time-decaying skill library that is re-injected next round, an Advisor decides whether to proceed, refine or pivot, and a Writer and Reviewer produce a number-grounded draft and review it. Guards against reward hacking are the stated focus: scoring is two-phase and frozen so training code cannot fake the metric or see the validation set, edits are restricted by an allowlist, numbers in the write-up must resolve through a verified registry, and a deliberately cheating engine is bundled to prove the defences fire. Every agent claim that can be checked is scored against reality, so the loop learns which of its own agents to trust. The suite runs on CPU with a mock model and needs no API key or GPU.

https://github.com/renee-jia/scholar-loop

autonomous-agentsmulti-agentai-researchai-scientistllm-toolspytorchexperiment-automationpython