AI research workflows in Claude Code and Codex

This is our own R&D: a method built and developed on our own projects, not a client commission. The closest evidence we can show is that the website you are reading was made through this process—from collecting the old site, through the brief and technical plan, to code.

What it is

A set of multi-agent workflows run in Claude Code and Codex, where many agents work on one task with separate roles and control points.

• Parallel-agent distillation. Material is divided into topic clusters; each receives an agent using the same note template, so results are comparable.
• Draft panel and judge. Strategic direction is not created in one pass. Three independent drafts are written separately, then a separate judge selects and explains a direction in writing.
• Adversarial review. A final agent tries to disprove the result: finding contradictions, inconsistent numbers, and unsupported decisions. The reviewer is not the author of the scope under review.
• Hard gates between stages. The next stage cannot start until the prior one has a completed definition-of-done checklist.

What problem it solves

• One agent loses consistency in a long task. Dividing work into roles and clusters means no agent carries the entire scope alone.
• The result depends on the mood of a session. Gates and templates make completeness repeatable even when content differs.
• A model confirms the first hypothesis. Independent drafts and a judge force comparison of alternatives.
• “AI wrote something, but nobody knows whether it is true.” Adversarial review assigns someone the explicit duty to look for errors before the result moves on.

How we built it

There is no single program; there is a workflow, written roles, and state held in files. A coordinating agent assigns work, protects contracts, and integrates checked results. Delivery agents work in parallel on separate scopes, while a separate reviewer tries to disprove the result. Claude Code and Codex are tools in this arrangement; the role depends on the task, not the model brand.

The key decision is that what is in files is the truth, not what a session remembers. Distillation, brief, technical plan, and decision log live in the repository and pass through git. Every session begins by reading the saved state rather than recreating it from memory.

This site illustrates the whole process: 81 old-site URLs archived and distilled with a decision for each page; a three-draft panel resolved by a judge; then a brief, technical plan, target URL architecture, and implementation with independent review. The content you are reading was created in the same workflow—files, not an admin panel.

What it demonstrates

It demonstrates that we can design and run a working multi-agent system, not only talk about one. It is evidence for multi-agent systems: a process that was actually run and the artefacts left by it.

The method and role split are described in detail in how we build with AI, and the client-facing version in the process.