Org-behavior professor KarlMuth shares his recipe for designing research agents

KarlMuth—an organizational behavior instructor with a background spanning AI research, law, and economics, who studied competitive decision-making under uncertainty at LSE—followed the viral success of his academic writing thread (boosted in part by economist Alex Tabarrok's reshare) with a follow-up guide on building research agents, prompted by a flood of DMs asking how.

Confirmed

For the common complaint "I gave it a pile of materials but it didn't use them / didn't do what I expected," his first piece of advice is structure first: handing the model a stack of PDFs does not equal a structured knowledge base.
Concrete workflow: use two triage agents to convert each document into a file (CSV, JSONL, or similar), adding annotations, decorations, and classification tags beyond the fields so every data point is genuinely usable; he demonstrates with a processed NYT article whose local file contains both the original text and added context. The accompanying image is his tool Veers (anonymized screenshot).
Once the data is in order, define agent responsibilities using role theory from Rizzo, House, and Lirtzman (1970): good roles have both a center of gravity and boundaries.
Place agents on the autonomy spectrum proposed by Andrew Ng, and use Morrison's (1994) organizational citizenship behavior (OCB) model to anticipate where agents might take liberties, over-help, or go off track.
He also argues: agents usually fail not from too many constraints but from too-low expectations; drawing on Amy Chua's educational philosophy, you should spell out the agent's refusal conditions, stopping rules, and expected outputs rather than merely imposing constraints.

Why It Matters

This methodology systematically brings classic organizational behavior and decision theory (including his jury and collective decision-making research) into agentic design, giving researchers an actionable role–boundary–autonomy framework rather than scattered tips.

2026-09-03 ~ 2026-09-03 · 10 related posts

Primary sources

Use Org-Design Theory to Structure Files and Roles for Reliable Agents — KarlMuth · 2026-09-03 22:45 (UTC+08:00)
Role Theory Plus Autonomy Spectrum: A Framework for Keeping Agents on Task — KarlMuth · 2026-09-03 22:40 (UTC+08:00)
"I gave it stuff but it didn't use it": structured inputs are the fix, says KarlMuth — KarlMuth · 2026-09-03 22:39 (UTC+08:00)

Alex Tabarrok amplifies KarlMuth's AI academic-writing thread, prompting a follow-up agent research guide — KarlMuth · 2026-09-03
Multiagent design through the lens of juries and decision theory, per an LSE-trained researcher — KarlMuth · 2026-09-03
[source] "I gave it stuff but it didn't use it": structured inputs are the fix, says KarlMuth — KarlMuth · 2026-09-03
Triage agents turn documents into structured, annotated files — stop feeding LLMs a pile of PDFs — KarlMuth · 2026-09-03
Don't Dump Raw Text Into Agent Pipelines: Annotate and Categorize Your Documents — KarlMuth · 2026-09-03
Design Agent Roles Before Deployment: Every Role Needs an Emphasis Center and Boundaries — KarlMuth · 2026-09-03
[source] Role Theory Plus Autonomy Spectrum: A Framework for Keeping Agents on Task — KarlMuth · 2026-09-03
Placing Agents on an Autonomy Spectrum and Using OCB Theory to Predict Over-Helping — KarlMuth · 2026-09-03
[source] Use Org-Design Theory to Structure Files and Roles for Reliable Agents — KarlMuth · 2026-09-03
Agents fail from low expectations, not high constraints, argues KarlMuth — KarlMuth · 2026-09-03