DarwinX and the Evolved Harness Thesis: What Population-Based Natural Selection Over Agent Configurations Means for Your Codex CLI Skills, Profiles, and Hooks

Most practitioners treat their Codex CLI configuration — AGENTS.md, skills, hooks, named profiles — as static preferences: write once, tweak occasionally, forget. DarwinX, a new framework from Zhang et al. (arXiv:2608.07545, July 2026), demonstrates that treating these harness components as an evolvable population under selection pressure yields gains that rival or exceed switching to a stronger model ¹. The implications for how you manage your Codex CLI configuration stack are significant.

The Core Insight: Evolve the Harness, Not the Model

DarwinX freezes the base LLM entirely and evolves only the harness — the combination of prompts, skills, tools, memory, and control flow that wraps the model ¹. It maintains an archive of harness variants organised as a tree, where each node stores edit deltas, per-task scores, trial evidence, and distilled lessons. Children are admitted only when they satisfy a preserve-and-extend contract: net improvement on at least one task (g(c) > 0) and bounded regression on others (R(c) ≤ δ) ¹.

This mirrors what Codex CLI practitioners already do informally — adjusting AGENTS.md instructions, adding skills, tuning hooks — but DarwinX formalises the process with population-based search and rigorous fitness evaluation.

flowchart TD
    A[Frozen LLM<br/>GPT-5.5 / Opus 4.8] --> B[Harness Population]
    B --> C[Variant 1<br/>Skills + Prompts]
    B --> D[Variant 2<br/>Tools + Control Flow]
    B --> E[Variant N<br/>Memory + Verification]
    C --> F{Preserve-and-Extend<br/>Contract}
    D --> F
    E --> F
    F -->|Pass| G[Archive &<br/>Steering Set]
    F -->|Fail| H[Revert]
    G --> I[Next Generation<br/>Parent Selection]
    I --> B

What Gets Evolved

DarwinX modifies two layers of the harness ¹:

Skill layer: prompts, memory structures, and distilled knowledge — analogous to AGENTS.md instructions, SKILL.md files, and Memories in Codex CLI
Code layer: tools, control flow, and the agent loop implementation — analogous to hooks (PreToolUse/PostToolUse), MCP server selection, and sandbox configuration

Three learning signals drive evolution:

Failure-derived (∇): failed trajectories are summarised to localise missing capabilities
Teacher-derived (π*): reference solver trajectories are distilled into reusable approaches
Self-derived (A): the agent’s own passing versus failing rollouts reveal reliable success patterns

The Numbers That Matter

DarwinX was evaluated across four benchmarks, each testing a different generalisation dimension ¹:

Terminal-Bench 2.1 (In-Domain)

On 89 terminal tasks with GPT-5.5 frozen throughout, DarwinX achieved 83.2% avg@5 — a +7.7 point gain over the base Monet harness (75.5%) ¹. The largest improvements appeared in ML/scientific-computing tasks (+14.8 points) and data/database tasks (+13.8 points). For context, the Codex CLI leaderboard entry on Terminal-Bench 2.1 sits at 83.1% with GPT-5.5 ¹.

WebArena-Infinity (Synthetic-to-Real Transfer)

Evolution used 300 synthetic intents scored by an LLM judge — the system never read the benchmark task suites. On 1,260 real tasks across 10 applications, audit-clean pass@1 jumped from 43.5% to 93.0% (+49.5 points) ¹. The validity audit is particularly striking: base invalid trajectories dropped from 293 (23.5%) to 17 (1.4%), with all evaluation-plane exploits and privileged-host mechanisms eliminated.

SWE-bench Verified (Cross-Benchmark Transfer)

A Terminal-Bench-evolved harness transferred unchanged to 500 SWE-bench Verified issues, achieving 84.2% pass@1 on Opus 4.8 — a +3.4 point lift over the fix-skill reference (80.8%) ¹. No in-domain evolution was conducted; the harness generalised purely from terminal task selection pressure.

TerminalWorld (Held-Out Generalisation)

Evolution on 94 training tasks reached 100% solve rate, but the 41-task held-out split achieved 68.3% pass@1 ¹. The 31.7-point gap reveals a proxy overfitting risk that Codex CLI users should recognise: optimising AGENTS.md for your current test suite does not guarantee generalisation to new tasks.

What DarwinX Evolved: The Verification-Contract Pattern

The seven skills that DarwinX’s evolution converged on all belong to what the authors call the verification/artifact-contract family ¹:

verifier-contract / contract-candidate
graded-artifact-final-check / artifact-verification-loop
real-tool-artifact / tool-grounded-artifact
security-contract-repair

The pattern is clear: the evolutionary process independently discovered that explicit acceptance contracts and tool-grounded output verification before finalising results are the highest-value harness additions. Extra compute on newly solved tasks doubled turns (22 vs. 11) and quadrupled tokens (380K vs. 89K), but compute on already-solved tasks barely moved ¹.

Mapping DarwinX to the Codex CLI Configuration Stack

The Codex CLI v0.147.0 configuration stack already provides the building blocks that DarwinX evolves over ²³:

DarwinX Component	Codex CLI Equivalent	Configuration Location
Skill layer (prompts)	AGENTS.md instructions	`<repo>/.codex/AGENTS.md` or `~/.codex/AGENTS.md`
Skill layer (memory)	Memories system	`~/.codex/memories/`
Skill layer (knowledge)	SKILL.md files	`~/.agents/skills/<name>/SKILL.md`
Code layer (tools)	MCP servers, enabled_tools	`config.toml` `[mcp_servers]`
Code layer (control flow)	PreToolUse / PostToolUse hooks	`config.toml` `[[hooks.PreToolUse]]`
Code layer (verification)	PostToolUse validation hooks	`<repo>/.codex/hooks.json`
Population management	Named profiles	`~/.codex/<profile>.config.toml`
Fitness evaluation	OpenTelemetry export	`config.toml` `[otel]`

Treating Named Profiles as a Population

DarwinX’s archive of harness variants maps directly to Codex CLI’s named profile system ³. Each profile overlays the base configuration with different model settings, reasoning effort, and approval policies:

# ~/.codex/verification-heavy.config.toml
model = "gpt-5.5"
model_reasoning_effort = "xhigh"
approval_policy = "on-request"

# ~/.codex/speed-specialist.config.toml
model = "o4-mini"
model_reasoning_effort = "low"
approval_policy = "never"

Invoking codex --profile verification-heavy or codex --profile speed-specialist switches the entire harness in one argument ³. What DarwinX suggests is that you should maintain multiple profiles and measure which performs best on your actual task distribution, rather than settling on a single configuration.

Implementing the Verification-Contract Pattern

The evolved verification skills map to PostToolUse hooks in Codex CLI. A basic artifact-verification hook:

[[hooks.PostToolUse]]
matcher = "^(Write|Edit)$"

[[hooks.PostToolUse.hooks]]
type = "command"
command = '/usr/bin/python3 ~/.codex/scripts/verify-artifact.py'
timeout = 30
statusMessage = "Verifying artifact contract"

The DarwinX evidence suggests this kind of post-execution verification is the single highest-value harness addition you can make ¹.

The Proxy Overfitting Warning

DarwinX’s TerminalWorld results — 100% on training tasks, 68.3% on held-out — carry a direct warning for Codex CLI practitioners ¹. If you optimise your AGENTS.md and skills for a specific codebase or task type, you risk overfitting your harness. The archive and merge mechanisms in DarwinX address this by maintaining diversity: four specialist harnesses solved 24–27 of 41 held-out tasks individually, but the merged harness achieved 28 ¹.

The Codex CLI equivalent is maintaining task-specific profiles rather than cramming everything into a single AGENTS.md. Use project-level .codex/AGENTS.md for repository-specific instructions and keep ~/.codex/AGENTS.md lean and general-purpose ³.

flowchart LR
    subgraph "Global Harness (lean)"
        GA[~/.codex/AGENTS.md<br/>Core conventions]
        GS[~/.codex/config.toml<br/>Base settings]
    end
    subgraph "Specialist Profiles"
        P1[security-review.config.toml]
        P2[refactoring.config.toml]
        P3[greenfield.config.toml]
    end
    subgraph "Project Overrides"
        PA[.codex/AGENTS.md<br/>Repo-specific rules]
        PH[.codex/hooks.json<br/>Project hooks]
    end
    GA --> P1
    GA --> P2
    GA --> P3
    GS --> P1
    GS --> P2
    GS --> P3
    P1 --> PA
    P2 --> PA
    P3 --> PA
    PA --> PH

What Codex CLI Still Lacks

DarwinX highlights several capabilities that Codex CLI’s configuration stack does not yet support:

Automated fitness evaluation: DarwinX uses benchmark-specific verifiers to score harness variants. Codex CLI’s OpenTelemetry export ([otel] in config.toml) provides raw telemetry ³, but there is no built-in mechanism to automatically score configuration changes against a task suite and accept or revert them.
Preserve-and-extend enforcement: when you edit AGENTS.md or add a skill, there is no automated check that the change improves some tasks without regressing others. DarwinX’s contract-based admission policy prevents exactly this failure mode ¹.
Archive and lineage tracking: DarwinX maintains the full tree of harness variants with edit deltas. Codex CLI profiles are flat files with no lineage, branching, or merge semantics ³.
Cross-specialist merging: DarwinX’s merge operator combines complementary specialist harnesses under a union-coverage criterion ¹. Named profiles in Codex CLI are mutually exclusive — you cannot compose two profiles into one.
Failure-signal distillation: DarwinX automatically distils failed trajectories into harness improvements. Codex CLI captures JSONL traces ² but does not automatically extract actionable configuration changes from them.

Practical Takeaways

The DarwinX results suggest a concrete workflow for Codex CLI practitioners:

Version your harness: treat .codex/AGENTS.md, config.toml, hooks, and skills as code — commit them, review them, track changes
Maintain multiple profiles: create task-specific named profiles rather than one monolithic configuration
Measure before and after: use [otel] telemetry or manual benchmarking to quantify the effect of configuration changes
Add verification hooks first: the evolved harness converged on post-execution verification as the highest-value addition — implement PostToolUse hooks before investing in prompt engineering
Keep global config lean: resist the urge to grow ~/.codex/AGENTS.md into a comprehensive manual — the TerminalWorld overfitting gap shows that specificity costs generality

The +7.7 point gain on Terminal-Bench 2.1 from harness evolution alone — without changing the model — is the clearest evidence yet that your configuration is not a preference; it is a performance lever ¹.

Citations

Zhang, Y., Dai, Y., Tan, J., Yang, L., Mullur, R., Hoang, T., Hu, Z., Zhu, J., Mui, P., Savarese, S., Xu, R. & Chen, Z. (2026). “DarwinX: Evolving Agent Harnesses Through Natural Selection.” arXiv:2608.07545. https://arxiv.org/abs/2608.07545 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹² ↩¹³ ↩¹⁴ ↩¹⁵ ↩¹⁶ ↩¹⁷ ↩¹⁸
OpenAI. (2026). “Codex CLI Documentation.” https://developers.openai.com/codex/cli ↩ ↩²
OpenAI. (2026). “Codex CLI Advanced Configuration.” https://learn.chatgpt.com/docs/config-file/config-advanced ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶