EvoX Matches Codex Scores at Just $1.95 per Task – How a Chinese Team Achieved It

According to Deloitte’s 2026 report “Why 86% of AI Agent Pilots Never Reach Scale”, 86% of multi‑agent deployments collapse in real‑world settings, and recent Silicon Valley cases show costly token waste when agents clash while handling supply‑chain tasks.

To make a group of smart agents both obedient and efficient, the EvoMap team introduced EvoX , which abandons a single “master” controller in favor of a “swarm‑style self‑evolution” strategy: each agent works on its own atomic sub‑task and the system automatically aggregates the results, mimicking natural bee‑hive coordination.

The article contrasts two prevalent patterns: (1) a lone agent that processes an entire task but suffers context truncation on long problems, and (2) an “Agent Team” where a primary agent delegates to sub‑agents, only to suffer a “telephone‑game” loss as each layer re‑interprets the output. The real bottleneck, therefore, is not the number of agents but the quality of role‑based collaboration, multi‑round convergence, and emergent organization.

EvoX evaluates agents on three criteria: can the complex task be fully covered, can the results be reliably merged, and can the system evolve a more effective organizational structure.

Benchmark Experiment 1 tested three modes—swarm (EvoX), traditional Agent Team (Claude Code style), and monolithic—across six mainstream benchmarks, totaling 563 tasks (100 logical, 250 ordinary math, 63 competition math, 150 physics). The swarm mode doubled fault tolerance, lifted long‑task completion rates, and cut costs dramatically. Detailed analysis revealed that while 373 intermediate answers were correct, only 217 survived the final aggregation, a retention rate of 55.5%, highlighting the loss in “telephone‑style” pipelines.

The three advantages of the swarm approach are: (1) finer task granularity with clear atomic boundaries, (2) execution isolation preventing context interference, and (3) direct aggregation without a summarising master, preserving correct answers.

Benchmark Experiment 2 placed EvoX, Claude Code, and Codex on the Opus 4.8 model across 424 independent tasks from six benchmarks. EvoX matched Codex (338 passed tasks) and trailed Claude Code by only 20 tasks (Claude Code passed 358). Specific results include: AppWorld 87/90 (EvoX only one task behind the top score), GAIA 49/51 (tied for highest), SWE Multilingual 65/68/66 (EvoX close to peers), while Terminal‑Bench 2 and Cybench remain weak spots.

Cost analysis shows EvoX’s per‑task price at $6.10 on the standard price list—lowest among the three—and an actual cached cost of $1.95, substantially below Claude Code and Codex, proving that performance gains are not bought with unlimited token budgets.

In a second experiment on self‑organization, 24 identical agents accumulated “Gene” memory after each task. When the system randomly broke some connections and let agents re‑select partners, agents that considered task performance and accuracy chose more effective collaborators than those relying solely on social ties, demonstrating that the information exposed to agents shapes the emergent network.

The article concludes that EvoX is not a finished product but a beta‑stage system that already shows measurable product capabilities: task decomposition, parallel execution, preservation of intermediate correctness, and stable delivery in complex environments. However, further work is needed on benchmarks like Terminal‑Bench 2 and Cybench, and on fully automating self‑evolution beyond the current static architecture.

Overall, EvoX illustrates a shift from one‑off token‑exchange APIs toward a continuously improving agent infrastructure that learns from experience, refines task boundaries, and self‑organises without a central brain.