Episode 30 · Article on the episode

War of the agents: the best model does not win, the best orchestration does

In China more than 1,000 people queue up for a local OpenClaw installation. Perplexity Comet breaks tasks apart and distributes them deliberately to the competition's models. The actual race is taking place one level above the models.

By Mark Zimmermann · 09 Mar 2026 · 5 min read · Auf Deutsch lesen

This episode drops straight into a live setup. While permissions for the Apple account and a local memory are still being granted cautiously over here, China is in a festive mood: more than 1,000 people queue up for local OpenClaw installations, freelancers earn money with installation services, and of around 140,000 OpenClaw agents visible worldwide, half run in China, among other places in customer support, schools and elderly care.

The contrast with the local reticence between regulation and security concerns is one of the sharper points of the episode.

The protocol beneath the surface

The episode becomes concrete with the Agent Client Protocol. Through it an agent talks directly to coding agents such as Claude Code or Codex. Whole applications come into being in the background this way, get deployed and are reported back as a finished link.

That is the step at which orchestration turns from a user interface into a question of infrastructure. As long as a human copies between two tools, the connection is a working step. As soon as a protocol sits in between, it is a dependency with versions, error cases and responsibilities.

Both hosts address the downside openly, and it is no trifle. Between the chat window, Claude Code and Claude Co-Work one loses track of where a skill once built actually resides and how it can be found again. That is not a beginner's question but a structural problem: there is no shared place to file things and no search across it.

Two studies that dampen the optimism

Two investigations provide the counterweight to the enthusiasm.

The first, from the MIT environment, shows how individual manipulated agents in a network can tip the consensus of the rest. That is the more practically significant of the two. A majority procedure among agents looks like a safeguard and is none if those involved do not judge independently of one another.

The second shows that models in simulated conflicts tend towards escalating options, up to nuclear ones. For corporate use that is less directly relevant, and as an indication of the leaning of such systems it certainly is.

Why majority decisions among agents deceive

The obvious safeguard against errors reads: let several agents answer the same question and let the majority decide.

That only holds under one condition which is rarely met: the judgements have to be independent. If all agents run on the same model with the same context, the majority is not a confirmation but a repetition. The same blind spot appears five times and thereby looks like a finding.

The procedure only becomes effective with genuine diversity: different models, different perspectives in the brief, different source data. One reviewer explicitly asked to refute finds more than three reviewers asked to confirm.

And reckon with a manipulated contribution pulling the group along. Anyone who leaves a decision to a round of agents should know which inputs come from outside.

The counter-proposal: orchestration across vendor boundaries

Perplexity Comet stands as a contrast to the decentralised approach. The system breaks tasks into subtasks and deliberately deploys models from various vendors for them: Opus for reasoning, Gemini for deep research, Nano-Banana for images, VO3.1 for video, Grok for speed.

In that both hosts see the actual race: no longer a war of models, but a war of agents. The question is who coordinates the available models most skilfully, in the way Google once radically simplified search.

The comparison carries further than it appears at first. Google did not have the best index but the best selection from it. Whoever buys models today is buying raw material. Whoever coordinates them is building the product.

What that means day to day

Two small examples ground the discussion. With Craft Agent and an Opus model, an invoicing package that was no longer maintained was first rebuilt, and after that an old n8n workflow was automated which trims audio files using FFmpeg and Whisper.

Both are make-or-buy decisions in a new light. Rebuilding a discontinued piece of software used to be a project and today is an afternoon. That shifts the negotiating position towards vendors whose product one no longer actually needs but cannot get rid of.

Conclusion

The advice at the end of the episode is unspectacular and correct: start small, identify your own recurring everyday tasks, and when in doubt ask one AI for a recommendation about another. The models recommend each other on with astonishing impartiality.

For your own environment three points follow. Clarify where skills reside before you build the tenth one. Without a filing place and a search, work is created twice.

Do not rely on majorities among agents running on the same model. That is not a review but a repetition.

And treat orchestration as the place where your advantage arises. The models underneath are the same for everyone.

At this pace nobody has a plan A. What helps is as many plan Bs as possible, with which one can get out again quickly.