Agentic Design Digest

Optimizations & self-improving agents

Optimizations & self-improving agents

Key Questions

What are Harvey self-loops in legal agents?

Harvey uses self-loops via harness engineering for 40-88% self-improvement in legal tasks. It drives agent learning in production. This exemplifies optimizations in agentic systems.

What is AutoKernel for GPU optimization?

AutoKernel is an open-source framework applying autonomous agent loops to GPU kernel optimization for PyTorch models, running 300+ experiments. RightNow AI released it for infra traction. It supports local hacks like Ollama Qwen3.

How does Gemma 4 enable on-device agents?

Gemma 4, Google's open models under Apache 2.0, are purpose-built for advanced reasoning and agentic workflows on-device. Guides cover local setups with Paperclip AI, Hermes on M4 MacBook. It pairs with Nanocode JAX for edge optimization.

What is MemFactory?

MemFactory is a unified LLM agent memory framework, achieving 95% on LongMemEval via Mastra. It supports self-improving agents with lifelong multimodal memory like Omni-SimpleMem. CMU self-org and Sakana Scientist advance this.

What local setups use Ollama and Gemma 4?

Local Ollama Qwen3 hacks and Gemma 4+Paperclip enable open-source agents, workflows on M4 MacBooks. Developer guides benchmark local deployment. This grows hands-on traction for self-hosted agents.

How is the AI industry automating research?

Sakana Scientist and industry races automate AI research with self-improving agents. KernelEvolve optimizes Meta’s AI infra via autonomous agents. AutoKernel and Harvey harnesses exemplify this trend.

What is Cog-DRIFT?

Cog-DRIFT enables models to learn from zero-reward examples in RLVR for skill internalization. It supports SKILL0 and agent optimizations. This aids self-org in CMU works.

What are key self-improving agent advancements?

Advancements include Harvey self-loops, Mastra/MemFactory memory (95% LongMemEval), AutoKernel GPU (300+ exps), and Gemma 4 on-device. Local Ollama/Gemma hacks and Sakana/CMU self-org drive traction. Nanocode JAX enhances infra.

Harvey self-loops (legal harness); Local Ollama qwen3 hacks/Gemma 4+Paperclip on-device; AutoKernel GPU (300+ exps); Nanocode JAX; Mastra (95% LongMemEval)/MemFactory/Sakana Scientist/CMU self-org; hands-on local/infra traction growing.

Sources (18)
Updated Apr 8, 2026