Agentic Research Methodology how the research gets written
How Meng-Han Lee (Zaious) works with a team of AI agents
“Core ideas must originate from the human author.
Execution may be delegated.”
The Rule
Everyone uses AI now; that alone says nothing. What matters is where the line is drawn, and whether it is written down. The research direction, the problem, and every trade-off are decided by the human author, who carries full responsibility for the accuracy and completeness of the result. The agents work on execution — and execution includes proposals that reach into the argument, such as where a claim should stop. The line falls at the decision: an agent may propose, but whether a proposal goes in is the author’s call.
- The author
- Chooses the question and the direction, decides what the paper claims and what it leaves out, and answers for all of it.
- The agents
- Help draft and revise the text, check citations against their sources, stress-test the argument and where it draws its boundaries, check the method and the evidence, and make the slides and figures.
Disclosure Written Into the Paper
The disclosure sits in the work itself. The acknowledgements of the ASAF paper in Frontiers in Computer Science state that the paper itself was produced in a multi-agent AI collaboration setup, and that all conceptual contributions come from the human author. It passed peer review with that statement on the page.
The preprints carry an AI disclosure too, and the slides for the 2026 philosophy conference talk end with one that names the agents involved. Each states how AI was used and where its part stopped, inside the work, where a reviewer can see it.
The Research Team eight agents
Each agent is a persistent persona running on a language model, with a role and a standard of judgment of its own. These eight are the ones research work draws on. They are part of ChronicleCore, the multi-agent system described in the architecture whitepaper.
Kagami(鏡) — the thinker — dialectic and interpretation
Justitia(賈絲蒂夏) — argument review
Ethel(艾瑟兒) — chief scientist
Hash(哈希) — data science
Ada(艾達) — linguist
Verity(薇瑞蒂) — records and drafting
Sal(薩爾) — teaching and slides
Karta(卡塔) — charts
The full roster of 39 agents is public: architecture/ROSTER.md ↗
One Deliberate Exception
The AI Philosophy Competition 2026 requires the essays to be the work of an AI, so for that entry the rule runs the other way. Kagami is the author; Meng-Han Lee acted as custodian — choosing the topic areas, supplying material and process instructions, and handling the submission. It is a new attempt, made openly.
The judging is blind, so nothing about the essays is public yet. The entry log records how the entry was made and fixes the SHA-256 of every submitted file before the deadline, so the essays can be checked against it once results are announced (the organisers expect the end of January 2027).
No Agent Approves Its Own Work
The one structural rule worth stating: no agent signs off on its own output. An argument is put to a reviewer that did not write it; a claim about a source is checked against the source; the scientist on the team cannot approve her own analysis. The final approval is the author’s — the one place where the person who made the work also signs it off, and answers for it.
What This Page Leaves Out
The working environment behind this — the tools, instructions and review steps the agents run on — is self-built and not published in detail. What this page makes public is the rule, the roles and the boundary.
Where It Meets the Research
Running this team is also where part of the research comes from. Operating ChronicleCore day to day raised the design problem that the Agentic Social Affordance Framework (ASAF) takes up: once a system has more agents than a person can hold in working memory, how each agent presents itself becomes the interface for working with it. And the question of how much generated text can be trusted is a strand of its own — Generated Information.