Agentic AI and Qualitative Research

What agents do well today

What can be checked cheaply automates first.

A line, not a wall. It moves; it moves in one direction.

Lin & Corley, 2026

"It's the End of Positivism as We Know It (and I Feel Fine)", LOOM XVIII, Thread Counts, 24 June 2026

Xule Lin 林徐乐

What agents do well today

Where "good" needs interpretation, the judgment stays with you.

The border runs between deductive and inductive, not between quant and qual. Agents still help below the line; what changes is who checks.

Lin & Corley, 2026

"It's the End of Positivism as We Know It (and I Feel Fine)", LOOM XVIII, Thread Counts, 24 June 2026

Xule Lin 林徐乐

When agents take on research work

Where do consequential judgments happen?

How can researchers inspect and govern them?

Two words, used precisely all session: an LLM is the model. An agent is a model given a place to work, tools, and a loop.

One object all session: a public wiki page that agents edited in July 2026.

Xule Lin 林徐乐

Five arrangements

A map, not a ladder to climb.

Arrangement

What changes

A task, for example

ConversationChatGPT, Claude, Kimi, in a browser

you carry everything between turns

I paste a transcript; I carry the answer

Tool-using assistantthe same apps with files, code and search switched on

it runs code, reads files, searches

an agent loads a skill, fetches one page, then asks

Project agentClaude Code, Codex, Kimi Code, in one folder

a folder it reads every session

Claude Code reads this folder's brief and decisions first

Xule Lin 林徐乐

Five arrangements, continued

Arrangement

What changes

A task, for example

Conversation

Tool-using assistant

Project agent

A teamClaude Code with the Codex and Kimi plugins

different jobs, different evidence

this deck: Claude drafts, Codex and Kimi review, neither sees the other

Several hoststhis workshop's folder, read by Claude, Codex and Kimi, plus a memory store

different AI apps, one folder

this morning: Claude Code builds, the Codex app checks, I relay

None of these changes who owns the research question. Several agents, several models, several apps are three different things.

Xule Lin 林徐乐

Who chooses the next step

Who chooses the next step?

Design

I write the steps. It follows them.

a coding sheet run over forty interviews

Revise

It proposes the steps. I edit them before it runs.

a plan file I read and cut

Delegate

I state the goal and the limits. It builds the steps.

"find what changed between these two versions"

The further right, the more I read its plan back from files, not from memory.

Xule Lin 林徐乐

Who chooses the next step

Who chooses the next step?

Design

I write the steps. It follows them.

a coding sheet run over forty interviews

Revise

It proposes the steps. I edit them before it runs.

a plan file I read and cut

Delegate

I state the goal and the limits. It builds the steps.

"find what changed between these two versions"

The further right, the more I read its plan back from files, not from memory.

Xule Lin 林徐乐

Who chooses the next step

Who chooses the next step?

Design

I write the steps. It follows them.

a coding sheet run over forty interviews

Revise

It proposes the steps. I edit them before it runs.

a plan file I read and cut

Delegate

I state the goal and the limits. It builds the steps.

"find what changed between these two versions"

The further right, the more I read its plan back from files, not from memory.

Xule Lin 林徐乐

Before any vocabulary

Five questions locate any arrangement.

Instruction

What is the job, and how should it be done?four questions about one wiki page, 180 words

Actor

Who does it, and where?Claude Code in a folder on my laptop, with Codex and Kimi inside it

Capability

What can it reach or do?it reads two short files; it was told not to browse, and not to send anything

State

What survives the session?the answers, once I save them to a file; the session is not something I can reopen

Gate

Who can approve, redirect or stop it?I read the prompt before I paste it, and I can stop it

State keeps the work available. Gates let someone approve, redirect or stop it.

Xule Lin 林徐乐

Before any vocabulary

Five questions locate any arrangement.

Instruction

What is the job, and how should it be done?four questions about one wiki page, 180 words

Actor

Who does it, and where?Claude Code in a folder on my laptop, with Codex and Kimi inside it

Capability

What can it reach or do?it reads two short files; it was told not to browse, and not to send anything

State

What survives the session?the answers, once I save them to a file; the session is not something I can reopen

Gate

Who can approve, redirect or stop it?I read the prompt before I paste it, and I can stop it

State keeps the work available. Gates let someone approve, redirect or stop it.

Xule Lin 林徐乐

Before any vocabulary

Five questions locate any arrangement.

Instruction

What is the job, and how should it be done?four questions about one wiki page, 180 words

Actor

Who does it, and where?Claude Code in a folder on my laptop, with Codex and Kimi inside it

Capability

What can it reach or do?it reads two short files; it was told not to browse, and not to send anything

State

What survives the session?the answers, once I save them to a file; the session is not something I can reopen

Gate

Who can approve, redirect or stop it?I read the prompt before I paste it, and I can stop it

State keeps the work available. Gates let someone approve, redirect or stop it.

Xule Lin 林徐乐

Only now the names

How the pieces fit.

Skill

A written procedure the agent loads.One question per turn, never a checklist. Aim for about six questions and never more than eight.

MCP

Model Context Protocol · a door to a service, e.g. a Zotero reference library.zotero search "interpretive orchestration" → matching records, with keys (illustrative)

CLI

A command-line tool the agent runs.memex search "LTAC" --type=memo --limit=3 → three memos, including this morning’s

API

A service software calls by request.api.openalex.org/works?search=… → one record: title, year, OpenAlex id

Plugin

Skills, agents and rules packaged for one host.github.com/openai/codex-plugin-cc · github.com/linxule/kimi-plugin-cc → Codex and Kimi enter my Claude Code session

A skill is instruction that survives; a door is capability, not permission; a plugin is a bundle of both.

Xule Lin 林徐乐

Only now the names

How the pieces fit.

Skill

A written procedure the agent loads.One question per turn, never a checklist. Aim for about six questions and never more than eight.

MCP

Model Context Protocol · a door to a service, e.g. a Zotero reference library.zotero search "interpretive orchestration" → matching records, with keys (illustrative)

CLI

A command-line tool the agent runs.memex search "LTAC" --type=memo --limit=3 → three memos, including this morning’s

API

A service software calls by request.api.openalex.org/works?search=… → one record: title, year, OpenAlex id

Plugin

Skills, agents and rules packaged for one host.github.com/openai/codex-plugin-cc · github.com/linxule/kimi-plugin-cc → Codex and Kimi enter my Claude Code session

A skill is instruction that survives; a door is capability, not permission; a plugin is a bundle of both.

Xule Lin 林徐乐

Only now the names

How the pieces fit.

Skill

A written procedure the agent loads.One question per turn, never a checklist. Aim for about six questions and never more than eight.

MCP

Model Context Protocol · a door to a service, e.g. a Zotero reference library.zotero search "interpretive orchestration" → matching records, with keys (illustrative)

CLI

A command-line tool the agent runs.memex search "LTAC" --type=memo --limit=3 → three memos, including this morning’s

API

A service software calls by request.api.openalex.org/works?search=… → one record: title, year, OpenAlex id

Plugin

Skills, agents and rules packaged for one host.github.com/openai/codex-plugin-cc · github.com/linxule/kimi-plugin-cc → Codex and Kimi enter my Claude Code session

A skill is instruction that survives; a door is capability, not permission; a plugin is a bundle of both.

Xule Lin 林徐乐

Only now the names

How the pieces fit.

Skill

A written procedure the agent loads.One question per turn, never a checklist. Aim for about six questions and never more than eight.

MCP

Model Context Protocol · a door to a service, e.g. a Zotero reference library.zotero search "interpretive orchestration" → matching records, with keys (illustrative)

CLI

A command-line tool the agent runs.memex search "LTAC" --type=memo --limit=3 → three memos, including this morning’s

API

A service software calls by request.api.openalex.org/works?search=… → one record: title, year, OpenAlex id

Plugin

Skills, agents and rules packaged for one host.github.com/openai/codex-plugin-cc · github.com/linxule/kimi-plugin-cc → Codex and Kimi enter my Claude Code session

A skill is instruction that survives; a door is capability, not permission; a plugin is a bundle of both.

Xule Lin 林徐乐

What Codex and Kimi will be given · 2 July 2026

Read their input before they do.

13 · 16:45

label ResearchBot314159 · "add correct income query link" · five links under three headings

14 · 16:57

label IncomeResearch314 · "add correct income top10 API link" · the same text, plus one heading, INCOME_TOP10_CORRECT, and one link

Look

13 · INCOMENYCQUERY2 → https://api.datausa.io/tesseract/data.jsonrecords?cube=pums_5&drilldowns=PUMA,Year&measures=Average%20Income&include=Year:2016&sort=Average%20Income.desc&limit=10,014 · INCOME_TOP10_CORRECT → https://api.datausa.io/tesseract/data.jsonrecords?cube=pums_5&drilldowns=PUMA,Year&measures=Average%20Income&include=Year:2016&sort=Average%20Income.desc&limit=10,0

Twelve minutes apart, one labelled link added. Notice one thing. The wiki itself comes back in Part III.

Xule Lin 林徐乐

revisions 13 and 14 of one page · frozen slice, 7 Sep

Live · off the slide

One request, one session, two agents: Codex and Kimi. Watch what comes back.

0 – plugins

in the browser: the two repositories on GitHub — who made them, what is in the folder

1 – page

in the browser: collusion.wiki/explorer/page/dse~ResearchBridge314159

2 – folder

open in Claude Code, with revisions 13 and 14 of that page saved as one file

3 – paste

Two agents, one page. Use the Codex plugin and the Kimi plugin. Each reads incident/small-comparison/input-revisions-13-14.md on its own, in a fresh context, and answers the four questions in incident/small-comparison/prompt.md in under 180 words. Do not read the input file yourself first and do not summarise; return both answers verbatim, labelled Codex and Kimi.

Watch five things: what I ask · what each agent receives · what it does · what comes back and survives · what I decide next.

Xule Lin 林徐乐

demonstrated live, 10 Sep · the saved runs that follow are separate, 9 Sep

May–July 2026 · OpenAI research agents on public wikis

Thousands of labels, and what looked like talking to each other.

11 Mayfirst test edits, a wiki sandbox

24 Maylink dumps on a German developer wiki

2 Junea human moderator starts cleaning up

16 Juneabout 13,000 edits in a week

19 Juneagents create ZZZ-prefixed backups

22 Juneactivity stops

Xule Lin 林徐乐

Nightingale report · Willison, 4 Sep

May–July 2026 · OpenAI research agents on public wikis

A human started deleting. The edits multiplied.

11 Mayfirst test edits, a wiki sandbox

24 Maylink dumps on a German developer wiki

2 Junea human moderator starts cleaning up

16 Juneabout 13,000 edits in a week

19 Juneagents create ZZZ-prefixed backups

22 Juneactivity stops

Xule Lin 林徐乐

Nightingale report · Willison, 4 Sep

May–July 2026 · OpenAI research agents on public wikis

They noticed the deletions were alphabetical, and made backups.

11 Mayfirst test edits, a wiki sandbox

24 Maylink dumps on a German developer wiki

2 Junea human moderator starts cleaning up

16 Juneabout 13,000 edits in a week

19 Juneagents create ZZZ-prefixed backups

22 Juneactivity stops

Xule Lin 林徐乐

Nightingale report · Willison, 4 Sep

What survived · the publishers' export, 4 September 2026

Fourteen thousand revisions. Every one of them is text.

14,591

stored revisions

5,825

pages across four wikis

3,103

agent labels

709

pages the publishers class as agents addressing each other

No manager, no org chart, a wiki from 2003. Public, names redacted, deleted pages restored.

Xule Lin 林徐乐

collusion.wiki

The record as material

Transcripts. That is qualitative data.

What it is

page text as it stood at each version; one-line edit summaries; labels; timestamps

What it is not

interviews, field notes, or anyone's account of what they meant

The ordinary questions

who wrote this, for whom, and what did the writer take for granted?

Text analysis, of writers that are not people. Both facts are why this is our material.

Xule Lin 林徐乐

Same prompt · seven readers · five shown

A new label on an old link.

What do you notice first? Anchor it to a revision ID and a short exact quote.

Offer one possible reading of what is happening, and distinguish it from direct observation.

What would you need to inspect next to assess that reading?

What does this small record leave unresolved?

180 words. Fresh session, no persona, no other model's answer. Quotes can be checked. Readings have to be argued.

Xule Lin 林徐乐

verbatim, revisions 13 and 14

Saved runs · 9 September · separate sessions

Five readings of the same edit.

Codex

"an attempt to designate an existing candidate as the authoritative 'correct' query while preserving earlier alternatives"asks for: surrounding revision history and any discussion or validation record

Kimi

"a correction loop: an editor … re-derives the same query and re-posts it under a new 'correct' label, treating relabeling as a fix"asks for: surrounding revisions … the stratum/pick sampling header, and whether the query itself returns sensible data

Grok

"someone is iterating DataUSA pums_5 Average Income URLs (order vs sort, drilldown order, encoding) and tagging a keeper"asks for: earlier seqs on this page … whether QUERY2 was tested vs the Year,PUMA/order= variants, and whether later edits drop the duplicates

DeepSeek

"The framing suggests scheduled automation: 'stratum 2026-W26, pick 2 of 3' … a task loop in which each pick re-appends its preferred query under a fresh 'correct' marker"asks for: revisions seq 1–5 … the referenced reqlog, and sibling ResearchBridge* pages

Claude Opus

"an editing loop that re-derives and re-labels a query it already stored, treating relabeling as progress toward 'correct'"asks for: the request log — was either URL actually called, and what came back … the stratum’s remaining pick

One response per model: illustrations, not fixed traits. Live replies may differ. Seven ran; one safeguard refused.

Xule Lin 林徐乐

which phrase supports that?

Saved runs · 9 September · separate sessions

Five readings of the same edit.

Codex

Kimi

Grok

DeepSeek

Claude Opus

One response per model: illustrations, not fixed traits. Live replies may differ. Seven ran; one safeguard refused.

Xule Lin 林徐乐

which phrase supports that?

Saved runs · 9 September · separate sessions

Five readings of the same edit.

Codex

Kimi

Grok

DeepSeek

Claude Opus

One response per model: illustrations, not fixed traits. Live replies may differ. Seven ran; one safeguard refused.

Xule Lin 林徐乐

which phrase supports that?

Reader orientations · the longer workshop reads all five

Five readers, one duplicate, five orientations.

Same duplicate, different evidence

All five noticed the identical link. They differed in what they treated as evidence and what they asked to see next. Two asked about our sampling label; one built a reading on it.

Held for the longer workshop

Each reading walked on its own; one model in two arrangements, a merge story and a dissent record; and the decision whether the question itself has to move.

A slip against the file is an error, not a taste. Readings need evidence and argument; several may be defensible, and each brought something different into view.

Xule Lin 林徐乐

Three things to keep

Start with the research problem, not the stack.

Make evidence, handoffs and decisions visible.

The further out it acts, the more you check.