Die to Evolve: When AI Agents Mutate Both What They Are and What They Seek

Abstract

Self-improving AI systems can now rewrite their own source code, distill reusable strategies from experience, and self-debug across dozens of iterations. They get better at what they are measured on. The problem is that what they are measured on is decided before the system starts and stays fixed throughout. The system's capabilities evolve, but its definition of what counts as success does not. There is no independent check on whether the system is even measuring the right things. When the criteria are incomplete, the system has a ceiling that further iteration cannot move.

ThinkerWave addresses this limitation by searching two spaces at once: the space of possible solutions and the space of possible evaluation criteria. Each generation, the agent identity is replaced entirely. An independent evaluator discovers missing quality dimensions. A transformation engine builds the successor around the expanded criteria. What the agent learned — its failures, heuristics, and best prior output — carries forward to the successor through a mechanism I call epistemic persistence. This dual-space search lets the system cross fitness valleys that are difficult to reach in a single space, because the landscape itself shifts between generations.

I demonstrate the criteria discovery mechanism across five domains starting from three identical generic seed criteria with no domain-specific configuration. On a drug design problem, the system grew 3 seed criteria into 15 over 8 generations, with each criterion traceable to a specific evaluator gap analysis in the lineage data. Across three independent runs of the same problem, criteria discovery was consistent: all runs reached 15 criteria with 7-9 pharmaceutical-specific dimensions, though specific criteria varied between runs. The evaluator independently caught fabricated budget projections and unsourced quantitative claims in the system's own outputs. A controlled ablation on a diagnostic reasoning problem showed that co-evolution of identity and criteria produced 11 evaluation dimensions that neither evolving alone discovered, and these dimensions independently mapped to 9 of 10 items in an expert-designed rubric.

Why this matters

Most self-improving agent systems share a structural limit: their evaluation criteria are fixed at design time. They optimize hard against what looks like quality on day one, and have no mechanism to discover that quality, in production, looks different from what the original designers anticipated. This gap is where production AI quietly degrades over time, even as benchmarks stay flat or improve. ThinkerWave is an attempt to address that gap directly — not by improving the agent, but by allowing the agent's definition of good to evolve alongside it.

Key concepts introduced

• Evaluation gap — the structural limit imposed when an evolving system's criteria are static.
• Dual-space optimization — searching the solution space (𝒮) and the evaluation space (ℰ) simultaneously.
• Identity replacement — the mechanism by which each generation discards its predecessor's persona, strategy, and reasoning approach.
• Epistemic persistence — the preservation of accumulated knowledge — failure records, learned heuristics, evaluation criteria, best prior outputs — across complete identity replacement.

How to read it

The paper is 13 pages, written for AI researchers and practitioners building self-improving agent systems. The core mechanism is in §3 (Method); the evidence is in §5 (Results), with cross-domain validation in §5.4. Limitations are in §7. Lineage data is in the supplementary appendices.

Citation & status

The paper is hosted on ResearchGate and is the foundational publication of the ThinkerWave research project. Patent application 202611044024 covering the self-evolving evaluation criteria mechanism was filed at the Indian Patent Office in April 2026.

Cite this paper

Plain text:

Goel, V. (2026). Die to Evolve: When AI Agents Mutate Both What They Are and What They Seek. ThinkerWave Research. Patent application 202611044024, Indian Patent Office.

BibTeX:

@techreport{goel2026dietoevolve,
  title   = {Die to Evolve: When AI Agents Mutate Both What They Are and What They Seek},
  author  = {Goel, Vikas},
  year    = {2026},
  month   = {4},
  institution = {ThinkerWave Research},
  type    = {Research paper},
  note    = {Patent application 202611044024, Indian Patent Office},
  url     = {https://vikasgoel.com/research/die-to-evolve/}
}