Codex Becomes a Super-Individual, EvoX Builds a Swarm — Two Roads for Multi-Agent

Codex is everywhere. In April, OpenAI shipped computer use, in-app browsing, image generation, memory, and plugins into the macOS and Windows Codex app (OpenAI Blog, 2026-04-16). In July, the $150 Codex Micro keyboard sold out in 14 hours. The same week, a four-month Pro user described running thirty image-model comparison cases by handing the whole job to Codex, going to sleep, and waking up to a finished spreadsheet (袋鼠帝, 2026-07-27; ifanr, 2026-07-24).

One agent. Every workflow. One human at the joystick.

That is Codex's road.

A small Chinese team called EvoMap just published a counter-road. On the same model (Opus 4.8), across six benchmarks and 424 tasks, EvoX matched Codex at 338 solved tasks, behind Claude Code's 358 by less than five points. Per-task cost under cache-aware pricing: $1.95 — visibly the lowest of the three (AI科技评论, 2026-07-26).

The numbers should not be possible. The bottleneck of multi-agent has been "the relay" — and the relay is exactly what EvoX throws away.

Two roads, same destination, opposite bets on who is in charge.

The Hidden Failure Rate of Multi-Agent

Most agent systems still look like a tiny company. A planner agent reads the task, breaks it down, dispatches to workers, and aggregates the result into a final answer. Claude Code and Codex are both built this way. It is the natural shape of an LLM and it is also where things fall apart.

A Deloitte 2026 report titled Why 86% of AI Agent Pilots Never Reach Scale puts the failure number on the record (cited via the EvoX release; original report pending verification). The EvoX team went further and instrumented the relay itself.

Across 563 tasks run on the Agent Team pattern, workers solved 373 correctly mid-process. After the planner summarized and aggregated, only 217 survived. 166 correct answers vanished in the relay — a 55.5% retention rate.

The takeaway is not "the model isn't smart enough." The model already has the right answer in its hand. The relay between workers and planner is what loses it. Each retelling is a re-compression, and re-compression eats signal.

What EvoX Actually Does

EvoX's bet is the opposite of "one smart boss." It splits every task into atomic parts with a fixed handler and a fixed output slot. A deterministic program collects the answers. The planner does not re-read or re-tell — it never sees the answers in narrative form.

Three things change:

1. Splitting is finer. Atomic tasks have explicit owners and explicit output positions. Coverage becomes traceable.
2. Execution is isolated. Each agent only sees its local problem. The drift caused by accumulated context disappears.
3. Merge is non-narrative. A program reads the slots. No agent summarizes, compresses, or interprets.

This is not magic. It is engineering. And the engineering pays off on the same benchmarks that built Claude Code and Codex's reputations.

On AppWorld (90 tasks), EvoX scored 87 — only one short of the top, and nine ahead of Codex. On GAIA (51 tasks), EvoX tied Claude Code at 49 (AI科技评论, 2026-07-26).

The model didn't get smarter. The merge protocol got quieter.

The $1.95 Knife

Benchmark scores get the headlines. Cost decides who ships.

Under unified list pricing, EvoX averaged $6.10 per task — the lowest of three. Under cache-aware pricing, the same runs landed at $1.95 per task, visibly below Codex and within range of Claude Code. The team isn't burning tokens to win; it's spending less because the merge is mechanical instead of generative.

This is the part most agent demos ignore. A demo that scores 90% by burning $20 per task is not a product; it's a research budget with a UI. The product is the one whose per-task cost stays flat as tasks get longer, messier, and more dependent.

The Swarm That Picks Its Own Teammates

The harder part is not the benchmarks. Real work does not arrive as 563 clean problems. Parts depend on each other, conflict, fail, and duplicate.

EvoX ran a second experiment: 24 same-config agents, each with its own memory of what it has done well — the team calls this a Gene. After enough rounds, each agent knows its specialty and accuracy. Then they choose teammates.

Only social information → agents form cliques, protect friendships, repeat past patterns. Add task and accuracy feedback → agents start picking partners by who solves what, not who knows who. Information design becomes organization design. Show agents their ties and you get a village. Show them their results and you get a project team.

Closing the loop is the next experiment: exchange Genes, fall back on failure, replace a teammate without re-running the whole task. EvoX is still beta. But the shape of what it's building is clear: agents that get better at organizing themselves the more they work.

Two Roads, Same Destination

Codex's bet: one agent, every workflow, computer use to drive the browser, plugins to reach the desktop, a hardware remote for the human at the controls. The Codex Micro joystick exists because Codex thinks the human's job is to steer one very capable thing.

EvoX's bet: do not build one boss. Build many dumb workers and a deterministic merge. Let the system learn which workers pair well with which tasks. Let the human design the merge protocol, not the playbook.

They are not competing on model IQ. Opus 4.8 is the same model on both sides of this comparison. They are competing on system architecture.

For builders, the practical question is which road to pick before the demo, not after. A team that starts with single-agent + deep computer use and tries to bolt on a swarm mid-project tends to inherit both sets of failure modes. A team that starts with a swarm and tries to give it a personality tends to rebuild the boss they tried to avoid. The middle is where projects go to die.

The bottleneck of multi-agent was never the model. It is the merge protocol. Whoever designs the merge protocol owns the next platform.