Research
Short notes on agents, training and serving. Each one links to the work it builds on.
Working notes from the team: what we tried, the public papers behind it, and where the idea stops helping.
Agent swarms on a knowledge graph When many cheap walkers beat one careful one.Many cheap agents leaving evidence-weighted traces on a knowledge graph find long causal paths more often than one agent walking alone.Read it
All notes. Newest first.
- Agent swarms on a knowledge graph When many cheap walkers beat one careful one.Many cheap agents leaving evidence-weighted traces on a knowledge graph find long causal paths more often than one agent walking alone.
- Environment mining for agent gyms Reviewed runs become test cases.Turning reviewed runs into gym environments: task, tools, frozen snapshot and a checker, with de-duplication, leakage guards and difficulty.
- Multi-LoRA serving and routing Many adapters, one base model.Many adapters on one base model: heterogeneous batching, paged adapter memory, a quality-and-cost router, and per-adapter evaluation.
- Recursion in agent delegation Depth, fan-out and when to stop.Delegation is recursion. Depth and fan-out set the cost, hand-offs compound errors, and shrinking budgets stop cascades.