🐙AI Agents Weekly: AlphaEvolve, codex-1, SWE-1, AI Agents vs. Agentic AI, OpenMemory MCP

AlphaEvolve, codex-1, SWE-1, AI Agents vs. Agentic AI, OpenMemory MCP

In today’s issue:

• Google announced a new coding agent for science

• OpenAI releases new SWE agent, codex-1

• Windsurf releases new SWE model

• OpenMemory MCP, a private memory for MCP clients

• New Agent-User Interaction Protocol

• Prime Intellect announces INTELLECT-2

• Flowise AI announces Flowise 3.0

• RL pipeline for training long-horizon language agents

• Top AI devs news, research, and much more.

---

---

Top Stories

AlphaEvolve

Google DeepMind introduces AlphaEvolve, a Gemini-powered coding agent designed for general-purpose algorithm discovery and optimization. It combines the creative synthesis of LLMs (Gemini Flash and Pro) with automated evaluation in an evolutionary loop to iteratively improve code solutions for complex problems in computing and mathematics.

• AlphaEvolve has been deployed across Google’s ecosystem, improving data center scheduling (recovering 0.7% of compute), accelerating TPU chip design through optimized Verilog rewrites, and boosting Gemini training speed by 1% through improved matrix multiplication routines. It also achieved a 32.5% speedup in low-level GPU kernel optimization (FlashAttention), significantly reducing engineering time.

• Beyond infrastructure, AlphaEvolve has made mathematical advances, including discovering a 4×4 complex matrix multiplication algorithm using only 48 scalar multiplications, improving over Strassen’s classical result. It also achieved progress on open problems like the 11D kissing number.

• The system flexibly applies to any problem where solutions can be encoded in code and verified. In a test of over 50 open math problems, it matched SOTA in ~75% and surpassed it in 20% of cases.

• AlphaEvolve represents a shift from symbolic or neural search-only approaches to a hybrid framework where LLMs not only generate code but also evolve and optimize full algorithmic solutions with quantifiable gains across real-world systems. An early access program for academic users is planned.

Blog | Paper