Definition: Recursive self-improvement (RSI) describes a system that improves its own capacity to improve, so each gain makes the next gain easier. The term covers a wide range in practice, from an agent rewriting one component of its own scaffolding to the theoretical case of a system redesigning its own intelligence without limit.
TL;DR: Most 2026 systems labelled "recursive self-improvement" rewrite their own scaffolding — the prompts, tools or search strategy around a fixed model — not the model itself. That distinction decides what the system can and cannot do, and most coverage blurs it. See a worked example.
Recursive Self-Improvement Explained in 3 Levels
Level 1, the everyday version. A person who gets better at learning, not just better at one subject. Each thing they pick up makes the next thing faster to pick up.
Level 2, the builder's version. An agent that notices which of its own habits waste steps, rewrites that part of its own setup, and keeps the version that measurably did better.
Level 3, the architect's version. A system that optimizes a component of its own optimization process. The critical question is which layer is being rewritten, because that determines whether gains compound or plateau at the fixed model's ceiling.
The Distinction That Decides Everything
When a paper or a product claims self-improvement, ask one question: what exactly gets rewritten?
| Layer rewritten | What changes | Compounds without bound? |
|---|---|---|
| The scaffolding | Prompts, tools, search strategy, orchestration code | No — bounded by the fixed model's ability |
| The context | Accumulated memory, playbooks, retrieved experience | No — bounded by what fits and what is relevant |
| The weights | The model's own parameters | In principle, but no deployed system does this in a loop |
| The architecture | How the model is built at all | Theoretical |
Almost every system described as recursively self-improving in 2026 operates in the first two rows. That is genuinely useful and worth building. It is not the third or fourth row, and treating them as the same thing produces bad predictions.
What Current Systems Actually Do
| System | Year | What it rewrites |
|---|---|---|
| AlphaEvolve | 2025 | Candidate programs, evolved against an automated evaluator |
| ShinkaEvolve | 2025 | Candidate programs, with far fewer samples |
| Darwin Gödel Machine | 2025 | Its own agent codebase, keeping only empirically validated changes |
| ReasoningBank | 2025 | Its memory, distilling past runs into reusable strategies |
| Agentic Context Engineering | 2025 | Its context, grown as an incrementally updated playbook |
| Dream-RSI | 2026 | Its exploration strategy, scored offline against recorded history |
Note what is absent from every row: the model weights. In Dream-RSI's case the paper is explicit that the coding agent, the evaluator and the model all stay fixed, and only the exploration-policy code is revised.
Why the Loop Usually Cannot Regress
A well-designed self-improvement loop includes the incumbent in the candidate set it compares against.
If the next version is chosen as the best-scoring among all candidates, and the current version is one of those candidates, then the selected version scores at least as well as the current one on that benchmark. This is a guarantee about the offline score on a fixed evaluation, not a promise about future behavior — but it is a cheap design property and worth copying into any tuning loop.
The Honest Limits
- Evaluation is the ceiling. A loop can only improve what it can measure. If the evaluator is wrong, the loop optimizes toward the wrong thing faster than a human would.
- Measurement noise caps the gain. If run-to-run variance exceeds the improvement being chased, the loop selects on noise. See evaluation noise floor.
- Bounded by the fixed model. Scaffolding improvements plateau at what the underlying model can do.
- Bounded by recorded ground. A loop that learns from its own history cannot learn about branches its history never touched.
Related Concepts
- Replay Simulator · World Model · Off-Policy Evaluation
- Evaluation Noise Floor · Evals · LLM as a Judge
- Agent Harness · Agentic Learning Loop · Reflection Pattern
- Reinforcement Learning · Scaling Laws · Emergent Behavior
Taskade does not run an automated self-improvement loop. What it does provide is the precondition: projects that record what was decided, agents grounded on them, and automations wired to both. See the app gallery for what that produces.
Read next: Replay Simulators Explained — a 2026 system examined closely, including the argument over whether it deserves the name.