Agent Harness
Studying how context, tools, memory, permissions, evaluation, and runtime shape agent reliability.
ResearchAgent Engineering · AI for Math · Education SaaS
Researching and Building Reliable Agent Systems
I research and deconstruct agent engineering while exploring practical applications of AI in education. My work focuses on Agent Harness engineering, LLM applications, AI for Math, and education SaaS productization.
Core Principles
- Middle structures determine reasoning quality
- System capabilities over standalone tools
- Verifiability over generative fluency
- Long-term consistency over short-term trends
LLM Application Engineering
Examining agent architecture, observability, cost, and engineering boundaries from prototype to production.
EngineeringAI for Math × Education SaaS
Applying formal reasoning and agent engineering to math learning systems and education products.
ApplicationCore Products
Dino-GSP (大角几何)
Dino-GSP (大角几何) is a dynamic geometry board and math AI tool for AI geometry and math education. It represents figures as computable structures, so construction, modification, and verification can be executed and reproduced.
Stage: many new features are being actively developed.
View product introductionAI Weekly Intelligence
A weekly intelligence product that curates AI updates and routes readers to product and research actions.
Weekly publication cadence running.
View product briefRecent Essays
AI for Math · Math Agent · Multi Agent · Formal Reasoning · Proof Verification
Why Mathematical Problems Are Hard for Agents in Practice
Cases from First Proof, Momus, LeanMarathon, Danus, and Aletheia reveal the central challenges math agents face in strategy selection, dependency management, long proofs, and preserving information from failed attempts.
17 min read
ReadLLM Engineering · Agent Efficiency · Context Engineering · LLM Cost · Model Routing
An Engineering Guide to LLM Token and Cost Optimization
A practical cost-optimization guide for agents, skills, MCP, AI coding, and model APIs, covering task economics, context, caching, tools, model selection, and stop conditions.
13 min read
ReadWork With Me
- Geometry engine integration
- AI content workflow systemization