Self-improving AI / PenguinHarness
PenguinHarness
“PenguinHarness is an open-source framework for agent self-evolution (recursive self-improvement, RSI). The model stays fixed; the harness evolves: prompts, Skills, tools and hooks.”
PenguinHarness is one of the projects on the self-improving AI map. Below is what it changes about itself, whether the change lasts, and whether it improves the improver, each with the line from its own source.
Where it sits
Columns: what it changes, from the surface to the deepest change. Rows: whether the change outlasts the run. A number is how many projects sit in that square. See it on the 3D map
- What it changes
- Tools, harness, its own code“The model stays fixed; the harness evolves: prompts, Skills, tools and hooks.”
- Does the change last
- Gone when the run ends“Every candidate is a new version, snapshotted before it changes, and every score lands in the Evaluation Center.”
- Does it improve the improver
- improves the task agent; no evidence the improver got better“Pick an algorithm, pick a Benchmark, and an agent improves another agent round by round.”
- Who sets the goals
- people set tasks and scoring“Pick an algorithm, pick a Benchmark”
- How it was tested
- demos only“a decision agent is evaluated, revises its working method once, and is retested on the same model. Its score goes from 84 to 100”
Date is when the GitHub repository was created. Yaowei Zheng, a maintainer, is listed with LlamaFactory and PrismShadow. The RSI algorithms (OPRO, APE, ACE, AWM and a default one) ship as plugins: an agent improves another agent, and the README says the harness (prompts, Skills, tools and hooks) evolves while the model stays fixed. Agents can also edit their own Skills; no case is reported of the improver being changed and then used, so recursion is task. Persists is run: versions are snapshotted, but no later use of an evolved version is reported. Measured is demo: the only self-evolution result is a video of one revision, retested on the same model, with no Benchmark named. The README's other numbers compare harnesses, not versions before and after evolution. It lists Train/Test pairs for six GDPEVO domains but reports no self-evolution results on their Test splits.
Does it count as RSI?
It depends on the definition. Under the four used on What counts as RSI, by what its own source says:
- Loose (Eliezer Yudkowsky, Intelligence Explosion Microeconomics (2013)): does not count
- Durable (Ren et al., self-improvement survey (2026)): does not count
- Strict (Roman Yampolskiy, From Seed AI to Technological Singularity (2015)): does not count
- Autonomous (Anthropic, When AI builds itself): does not count