Coding agents evolve without running candidates: Elite lifts SWE-bench Pro from 60.0% to 85.6% with zero candidate task evaluations
Synopsis
Self-improving coding agents usually run candidate versions on downstream tasks and let scores decide which lineage continues, so wider exploration raises evaluation cost. Elite instead implements several focused changes independently from the same parent, inspects the mechanisms in those executable programs, and reimplements a compatible subset as one descendant, using only deterministic structural and operational checks inside the loop. On a 60-task SWE-bench Pro subset, Elite raises the initial agent from 60.0% to 85.6% with zero candidate task evaluations, cutting observed evolution wall-clock time by 8.8x and 3.2x relative to DGM and HGM.
Interpretation
Recursive self-improvement of coding agents can proceed without scoring candidate versions on downstream tasks, because the evidence for what to inherit moves from task outcomes to the executable programs themselves. Prior systems evaluate and select: the Darwin Gödel Machine uses benchmark performance to pick parents from an archive, and the Huxley-Gödel Machine estimates a lineage's potential from descendant task performance, so broader exploration requires more task executions. On the 60-task SWE-bench Pro subset, Elite raises the initial agent from 60.0% to 85.6%, against 86.7% for DGM and 88.3% for HGM, while using zero candidate task evaluations.
Each generation proposes a set of focused improvement hypotheses grounded in concrete limitations of the current agent's runtime code, implements each independently from the same parent as an executable child, and then reads those implementations as evidence. The integrator judges mechanisms along five dimensions, complementary, redundant, conflicting, incomplete and overly complex, rather than assigning each child a predicted score or keeping one child whole. An integrated descendant is accepted only after smoke tests of the tool-call protocol and main execution flow, otherwise the current parent is retained; in one DeepSeek-V4-Flash run the exploration gate accepted 41 of 48 candidates and the acceptance gate accepted 5 of 7 integrated descendants.
Dropping candidate task evaluation also changes the cost structure of evolution, because the decision cost no longer grows with the size of a downstream evaluation set. DGM and HGM performed 810 and 800 candidate task evaluations during evolution, respectively, while Elite performed none. On DeepSeek-V4-Flash, Elite completed evolution in a mean of 4.90 hours versus 43.15 hours for DGM and 15.66 hours for HGM, observed speedups of 8.81x and 3.20x, with 16.35 self-modification attempts per hour against 0.58 and 3.77.
Perspective
The demonstrated scope is coding agents and selected benchmark subsets: a 60-task SWE-bench Pro subset for the evolution comparison, 50-task SWE-bench Verified and Polyglot subsets used only for transfer, plus ten evolution rounds on mini-swe-agent v2.4.6 with a fixed Claude-opus-5.5 backbone. The intended readers are researchers and engineers working on agent self-improvement and scaffold design.
Evaluation focuses on coding tasks and selected subsets, with limited repetitions for some comparisons; downstream performance fluctuates across accepted updates, making longer-term evolution an open subject; the current implementation maintains a single lineage, and the authors propose an archive of historical agents to preserve alternative evolutionary paths as future work.
