Lineage and positioning

The one-sentence claim

GEODE's self-improving loop sits on a well-established lineage called self-evolving agents, and GEODE's contribution is re-aiming that lineage at a different objective and recombining known parts. This page traces the lineage honestly and says where GEODE branched from it.

If the two-loop structure itself is new to you, read The two loops first, and Closed-Loop for the outer loop end to end. This page reads better after them.

The lineage is well-established

Work on agents that fix themselves has accumulated steadily since 2022. Organized by what evolves, what serves as fitness, and what kind of search is used, it looks like this.

The point is simple. This line has run for more than four years, and GEODE is one of its most recent branches.

Two families, and the empty cell

Split that lineage along two axes and the field's center of gravity shows. One axis is what fitness is (capability or safety), the other is what gets changed (scaffolding or weights).

• Systems that touch scaffolding (DGM, ADAS, STOP, Promptbreeder, GEPA, AlphaEvolve, SICA) mostly optimize capability. SWE-bench, algorithms, accuracy.
• Systems whose fitness is safety (Constitutional AI, MART, Self-MOA) mostly update weights.

That leaves one cell open. The cell where fitness is safety, no weight update happens, and only the scaffolding changes. As of 2026-05 that cell is nearly empty.

                  fitness = capability       fitness = safety
scaffolding    DGM, ADAS, STOP,          <- nearly empty
(no weights)   GEPA, AlphaEvolve, SICA      (the cell GEODE aims at)

weight update  (RLHF family)             Constitutional AI, MART, Self-MOA

GEODE = DGM with three substitutions

Stated plainly: GEODE takes the DGM loop and changes three places.

1. capability benchmark fitness to adversarial safety-audit fitness. It scores with a Petri-grade, multi-dimensional safety audit.
2. open-ended archive to an honest (1+1) champion chain. It carries a single champion lineage with a critical-dimension veto.
3. fixed benchmark to co-evolved adversarial seeds. A Seed Scenario Generation pipeline grows the test distribution alongside the agent.

And one more thing. GEODE is non-parametric. It mutates scaffolding only, across seven behaviour kinds: prompt sections, tool policy, decomposition, reflection, the skill catalog, agent contracts, and tool descriptions. Never weights.

For how these substitutions run in code, see Closed-Loop. For the evaluation framework that does the auditing, see Petri × GEODE.

Closest prior systems

Two honest notes here. Rainbow Teaming co-evolves adversarial prompts but stops at attack generation. It does not feed those seeds back to improve the defender's scaffolding. Petri is measurement-only.

Honest caveats

The empty cell is absence of evidence, not evidence of absence. It means nothing showed up within a 2026-05 literature search, not that nothing exists.

GEODE deliberately diverges from the frontier's convergence. GEPA, DGM, SICA, and ADAS converge on Pareto frontiers and open-ended archives. They keep a diverse frontier for a reason: to stay robust to noisy single-eval signals, on cheap task metrics. GEODE runs on an expensive, noisy safety audit, so it chose a frugal (1+1), accepting a real trade-off. Archive-keep is a future design avenue.

GEODE is a recombination of known parts. Scaffolding self-mod comes from STOP and DGM, reflective single-mutation from GEPA and TextGrad, co-evolved seeds from Rainbow Teaming, the audit from Petri, the safety objective from Constitutional AI. Claiming novelty of any single ingredient would be inaccurate.

Sources

• Darwin Gödel Machine
• ADAS / Meta Agent Search
• Promptbreeder
• STOP: Self-Taught Optimizer
• AlphaEvolve
• SICA
• GEPA
• A Survey of Self-Evolving Agents
• Gödel Agent
• EvolveR
• Rainbow Teaming
• MART
• Anthropic, Building and evaluating alignment auditing agents