Summary (Overview)

• DarwinX is a system that improves LLM agents by evolving their harness (prompts, tools, skills, control flow) through natural selection over a population of variants, while keeping the base model frozen.
• The core innovation is a preserve-and-extend contract: a child harness variant is admitted only if it improves on at least one task without regressing more than a small tolerance on tasks its parent solved, preventing cross-task interference.
• DarwinX maintains an archive of alternative lineages, allowing complementary specialists to be recombined via a merge operator, overcoming the path-dependence and early-edit plateau problems of single-lineage self-editors.
• Across four benchmarks with increasing separation between evolution signal and test, DarwinX adds ~17 points on average: Terminal-Bench 2.1 (+7.7 to 83.2%), TerminalWorld held-out (to 68.3%), WebArena-Infinity audit-clean (43.5% → 93.0%), and cross-benchmark transfer to SWE-bench Verified (84.2%).
• Selection is driven purely by measured fitness (avg@k solve rate under each benchmark's own verifier)—no gold solutions, no hand-picked winners—demonstrating that a frozen model need not be a fixed agent.

---

Introduction and Theoretical Foundation

The paper addresses a fundamental question: an LLM agent's capability depends not only on model weights but on its harness—the prompts, tools, memory, and control flow that mediate the model. Recent work makes harnesses self-improving through an inner optimization loop: batch rollouts, reflect, propose a bounded edit, gate against a held-out signal.

Two failure modes motivate DarwinX's design:

1. Path dependence: Single-lineage self-editors are biased by early edits and plateau (as reported by Robeyns et al. in SICA).
2. Cross-task interference: An edit that fixes one task family silently regresses another, so evolution over a mixed task distribution stagnates.

The paper builds on the Darwin Gödel Machine (DGM) framework, which uses open-ended archives with stochastic parent selection. However, DGM scores a child only against its parent, so lineages solving complementary tasks are never recombined. DarwinX makes the preservation obligation explicit and adds recombination across lineages.

Key theoretical insight: The harness is the learnable surface when the model is frozen. Selection over harness variants converts evaluation compute into durable capability, treating self-evolution as a selection problem rather than a training problem.

---

Methodology

2.1 Fitness and the Preserve-and-Extend Contract

Each variant is scored by its per-task solve rate $\hat{p}_t(v)$ (avg@k). For a child $c$ and parent $p$, the per-task change is:

$$\Delta_t = \hat{p}_t(c) - \hat{p}_t(p)$$

summarized as a net gain:

$$g(c) = \sum_t \Delta_t$$

and bounded regression:

$$R(c) = \sum_t (-\Delta_t)_+$$

The fitness enabler admits a child that extends without breaking preservation, i.e., $g(c) > 0$ and $R(c) \leq \delta$. A reasoned verifier agent $f$ adjudicates in two stages (promote, then probe), reading trial evidence $E$ and shared memory $K_g$:

$$\text{verdict}(c) = f(g, R, \mathcal{E}, K_g) \in \{\text{promote}, \text{revert}\}$$

A promoted child is re-tested at higher fidelity with a preservation probe before it may steer search. Each node carries a lineage gain $G(c) = G(p) + g(c)$ used for parent selection.

2.2 Branch Evolution

Each round picks a promising harness from the archive and proposes one additive edit aimed at a task the agent currently fails. Parent selection samples:

$$p^* \sim (1-\beta)\delta_{\text{argmax}_{v \in S} G(v)} + \beta \cdot \text{Broaden}(P)$$

With probability $1-\beta$, it exploits the highest-gain node in the steering set $S$ of confirmed variants; otherwise it broadens across the population $P$.

2.3 Population and Recombination

Every scored variant is retained as an archive node. Variants are classified by how their solved set $S(c)$ compares to the parent's $S(p)$:

• Improvers ($S(c) \supseteq S(p)$) and neutral children ($S(c) = S(p)$) preserve inherited solves and stay eligible.
• Stepping stones ($S(c) \subsetneq S(p)$) and archived nodes (trade solves) contribute only distilled lessons.

For complementary variants $v_1, \ldots, v_n$, the merged harness is:

$$H = H_0 \oplus \Delta \quad \text{with} \quad \Delta = \Delta_{\text{code}} \oplus \Delta_{\text{skill}} \oplus \Delta_{\text{prompt}} \oplus \Delta_{\text{tool}}$$

The child is kept if it covers the union of parents' wins: $S(\text{child}) \supseteq \bigcup_i S(v_i)$.

2.4 Learning Signals

Three native signal types drive proposals:

• Failure-derived signals (∇): Summarize failed trajectories and localize missing capabilities.
• Teacher-derived signals (π*): Distill a reference solver's successful trajectory into a reusable approach.
• Self-derived signals (A): Contrast the agent's own passing and failing rollouts.

All signals are translated into harness edits; none update model weights.

---

Empirical Validation / Results

RQ1: In-Domain Test-Time Evolution (Terminal-Bench 2.1)

On frozen GPT-5.5, DarwinX lifts base Monet from 75.5% to 83.2% (+7.7 points). On GPT-5.6 Sol at medium effort, it reaches 84.7%, at the frontier of the verified leaderboard, surpassing Claude Code + Fable 5 (83.8% at xhigh) and OpenAI's native reference (81.8%).

Table 2: Terminal-Bench 2.1 avg@5 (verified leaderboard)

The gain is the harness, not compute: extra turns/tokens concentrate on newly-solved tasks (22 vs. 11 turns, 380K vs. 89K tokens), while compute on already-solved tasks barely moves.

RQ2: Held-Out Task Generalization (TerminalWorld)

On 41 disjoint held-out tasks with frozen Opus 4.8, DarwinX reaches 68.3% (28/41), best on the split and above every of-the-shelf agent. The matched base improves from 25 → 28 tasks (+7.3 points). Notably, the in-loop proxy overfits (training subset saturates from 0.505 to 1.000), yet the merged harness beats every individual specialist (which solve 24–27 tasks), demonstrating the value of archive diversity over greedy proxy-following.

RQ3: Synthetic-to-Real Generalization (WebArena-Infinity)

Evolution operates on 300 synthetic intents scored by an LLM judge; reporting uses deterministic pass@1 on 1,260 unseen real tasks. Audit-clean pass@1 rises from 43.5% → 93.0% (+49.5 points), outperforming GPT-5.5 + Browser Use (86.1%) by 6.9 points and the top public agent (69.3%) by 23.7 points.

Table 4: WebArena-Infinity per-application audit-clean pass@1 (excerpt)

Critically, capability and compliance improve together: invalid trajectories fall from 293 to 17, with evaluation-plane, privileged-host, and exploit mechanisms disappearing entirely. The evolved harness adds four contract-oriented browser skills (e.g., web_task_contract, browser_spa_state_contract).

RQ4: Cross-Benchmark Transfer (SWE-bench Verified)

A Terminal-Bench 2.1 harness run unchanged on all 500 SWE-bench Verified issues reaches 421/500 (84.2%) official pass@1, +3.4 points over the 80.8% fix-skill reference, without any SWE-V feedback.

RQ5: Ablation — Evolved Skill Bundle

The evolved lineage adds seven skills, all in one verification/artifact-contract family:

Gains concentrate where a frozen base has headroom: ML & scientific-computing (+14.8 points) and data/database (+13.8 points), while already-strong clusters stay within noise.

---

Theoretical and Practical Implications

1. Selection over population beats single-lineage search: The TerminalWorld result directly demonstrates that retaining diverse specialists and merging them outperforms greedily following the in-loop proxy score (which saturates at 1.000 while held-out truth is 68.3%).

2. Verification-before-finalization as a transferable mechanism: The same skill family (establish an acceptance contract, check against it before committing) emerges independently across terminal and browser modalities, suggesting a general principle for harness design.

3. Frozen models are not fixed agents: The harness is the surface that can still move. Selection over it converts evaluation compute into durable capability, making harness evolution an amortized asset across model generations.

4. Preservation as a policy interface: The bounded-regression contract can be read as a safety property—the probe set decides which behavior is protected, turning operator requirements into selection constraints rather than post-hoc filters.

5. Auditability: Harness edits are human-readable diffs, unlike weight updates, providing an oversight primitive for self-improving systems.

---

Conclusion

DarwinX improves a frozen-model agent by selecting over a population of harness variants rather than training new weights. Three parts compose it: a preserve-and-extend contract, an archive retaining alternative lineages for recombination, and a signal interface unifying failure-, teacher-, and self-derived evidence. Selection is driven only by measured fitness under each benchmark's verifier.

Key results: TB2.1 rises from 75.5% to 83.2% (GPT-5.5) and to a leaderboard-frontier 84.7% (GPT-5.6 Sol); TerminalWorld held-out reaches 68.3% (best on split); WAI audit-clean pass@1 rises from 43.5% to 93.0% as invalid trajectories fall from 293 to 17; a TB2.1 harness transfers unchanged to SWE-bench Verified at 84.2%.

Limitations: The experiments evaluate the complete system; the archive, parent selector, and merge operator are not independently randomized. The verification/contract mechanism is a plausible explanation rather than a per-skill causal estimate. TerminalWorld has only 41 held-out tasks (one solve = 2.4 points), making the one-task margin over Claude Code suggestive rather than decisive.

Future directions: Separating operator effects under controlled budgets, co-evolving model and harness, treating the harness as an asset across model generations, and evolving against refreshed proxy suites for deployment. The broader point: a frozen model is not a fixed agent—the harness is the surface that can still move, and selection over it converts evaluation compute into durable capability.