Kevin Gu is a co-founder of ThirdLayer, which builds context infrastructure and continual-learning tools for enterprise AI. His open-source AutoAgent project explores how an agent can improve another agent’s harness through repeated evaluation.[2][3]
Start with AutoAgent →
Career
Gu’s founder biography describes prior mathematics and statistics study at Harvard and quantitative or machine-learning research at Jump Trading, IBM Research, QuantCo and Meta. ThirdLayer’s current product description focuses on organizational context that agents can retrieve and learn from, including traces and workflow information.[2][4]
Contributions
AutoAgent separates the human’s program.md from the agent.py file a meta-agent edits. The former defines the task and optimization instructions; the latter contains the agent’s prompts, tools and orchestration. Each candidate is evaluated on a benchmark, then kept or discarded. A fixed Harbor adapter boundary keeps changes to the harness separate from the evaluation interface.[3]
Ideas
The repository treats improvement as a series of scored experiments rather than a single prompt rewrite. A human chooses the task, evaluation and resource limits; the meta-agent searches within that setup. The resulting score therefore depends on the chosen benchmark and budget. It does not establish that every edit will transfer to a different deployment.[3]
Tools, writing & talks
Tools
Start here
AutoAgent
github.comAn evaluation loop that edits agent harnesses while preserving the benchmark adapter boundary.
Read program.md and the fixed adapter contract first. Running experiments requires the documented Harbor environment, model credentials and a task-specific compute budget.
- AutoAgentgithub.com
Meta-agent that autonomously iterates on agent harnesses overnight, tuning prompts and tools against benchmark scores.
Catalog workOriginal source
Writing, research & talks
An open source library that uses a meta-agent to autonomously improve task agents by experimenting with prompts, tools, and orchestration until performance climbs. Achieved #1 scores on SpreadsheetBench and TerminalBench through fully automated optimization over 24+ hours.
Catalog workOriginal source
x.com