Self-evolving agents went from papers to production this year. Here is the full stack for building an agent that gets smarter every week, with the verified numbers behind every layer
In May, DeepMind’s AlphaEvolve found a way to multiply 4x4 matrices in 48 scalar multiplications. The previous record stood for 56 years. The same system’s scheduling heuristic has been recovering 0.7% of Google’s worldwide compute for over a year, and its kernel fix cut Gemini’s training time by 1%.
One autonomous research system ran for 417 hours and produced 166 fully AI-generated papers for about $180K. Another built GPU kernels that beat expert baselines.
Meanwhile, your agent asked you the same clarifying question it asked on Tuesday.
That gap has a name now. AI researcher Shunyu Yao just published the taxonomy the field converged on, and it splits every self-improving system with one equation:
Agent = Model + Harness
The model is the weights. Improving it takes GPUs, RL pipelines, and a lab. The harness is everything around the weights: the prompts, the memory, the tools, the playbooks. Improving that takes markdown files and a loop, and it is where every result you just read about at the practical level actually lives.
Here is the free insight worth keeping: the harness layer is open to you today, and the measured gains are absurd. Reflective prompt evolution beat reinforcement learning by up to 20 points using 35x fewer rollouts. Evolving playbooks added +10.6% on agent benchmarks while cutting adaptation cost 83.6%. A memory layer cut tokens 90% and latency 91%. An agent that writes its own tools set a new GAIA state of the art while spending 15% fewer tokens. Zero of these touched model weights.
Most people rent intelligence and let every lesson evaporate at the end of the session. The teams pulling away run the same rented model inside a harness that compounds.
Behind the paywall, The Self-Evolving Agent Stack, the five layers with the copy-paste system for each:
▫️
the ACE-style evolving CLAUDE.md, with the weekly miner prompt and the two failure modes that kill most playbooksThe Playbook layer,▫️
the Mem0 pattern rebuilt with plain files, and the honest 6-point trade it makesThe Memory layer,▫️
the rule that turns any task done twice into a permanent tool, with the state-of-the-art numbers behind itThe Skill Factory,▫️
how skills compound into domain expert agents and a router, the Alita-G moveThe Specialist layer,▫️
GEPA’s 35x-cheaper-than-RL method as a manual workflow you run in an afternoonThe Prompt Evolution loop,▫️
playbook bloat, context collapse, skill sprawl, and the cross-domain confusion trap, each with its fixThe guardrails,▫️
from zero to a compounding agent in one week of normal workThe 7-day install plan,▫️ what the labs keep for themselves (weight-level learning) and the one paper that says harness and model evolve together nextThe frontier map,
One subscription unlocks every system
This is one build in a growing library. Premium opens all of them:
▫️ [The Prompting and Context Engineering library](https://www.the-ai-corner.com/t/prompting-and-context-engineering?r=1krivi)
▫️ [The Claude and Anthropic library](https://www.the-ai-corner.com/t/claude-and-anthropic?r=1krivi)
Plus 3 fresh systems every week. One playbook rule that saves one repeated correction pays the subscription back before next week.
The five layers, the copy-paste prompts, the verified numbers, the guardrails, and the 7-day install, in one system.
Get The Self-Evolving Agent Stack below 👇 #
Keep reading with a 7-day free trial #
Subscribe to The AI Corner to keep reading this post and get 7 days of free access to the full post archives.