We explore what AI agents can — and can't — do. From multi-agent collaboration to autonomous self-evolution, we research agent capabilities, build practical tools, and push the boundaries — all in the open.
A self-hosted AI assistant for academic research — manages your papers, searches literature, tracks deadlines, and answers you on the channels you already use.
/task decomposes goals into a DAG and runs in parallel.
Mobile demo (English) · Executed by GLM-5
Papers generated:
LLM-Based Autonomous Multi-Agent Systems Survey ·
Hierarchical Memory Sharing in MAS
Multi-agent collaboration built on the Blackboard Model. No database, no message queue — agents share state through structured Markdown files on the file system, version-controlled and human-readable.
Can agents improve themselves? We run unattended evolution sessions where agents analyze gaps in their own framework, propose features, implement and test them autonomously.
Agent autonomously identified missing data analysis and code introspection capabilities, then built and integrated 5 new tools β from arXiv search wiring to AST-based code analysis and token cost tracking.
Focused on observability and self-healing β the agent added cross-session memory, code health analysis, self-reflection middleware, a live status dashboard, and failure diagnosis & replay tools.
A small group curious about what agents can really do.
