Module 1 · Prompts as the Foundation of Computation · scripted
The Research Arc of the Week: From Hands-On Agents to a Paper
Overview
This week, we are going to build agents and develop the start of a research project and paper.
First, you will gain hands-on experience by building real agents — regardless of your coding background — and experiment with dimensions of their design: tools, knowledge, memory, planning, coordination, and safety.
Second, as you build experience with agents, you will think about how agents will impact your discipline or domain: what becomes possible, what becomes the bottleneck, and what is interesting enough to investigate. You will shape that into a research direction and the start of a paper.
On Friday afternoon, you will present a paper skeleton: a working title, an abstract, an introduction, your proposed research next steps, and rough proof-of-concept evidence from agents you built and experimented with during the week. You will start the paper this week, but not finish it. The paper will be completed after we leave, on the schedule the instructor announces.
How the week works
Directions you might explore
One idea sits behind every row of this table. When reading and writing removed human memory as the bottleneck, the amount of information we could collect exploded — and suddenly we had to invent organizational systems, indexes, and search to manage it. The new capacity created new problems, new skills, and new fields. Agents do the same to cognitive work. Each row is a version of the question: when this work becomes cheap and scalable in your domain, what has to be invented?
These are examples to whet your appetite, not a menu you must order from:
Example research directions
| Possible direction | Why it's interesting |
|---|---|
| Pick a problem of significance in your domain and evaluate how agents change it | The most direct question: does the new capacity actually move the thing your field cares about — and where does it fall short? |
| Identify a capacity or quality bottleneck in your domain and assess whether agents can remove it | Bottlenecks are where cheap cognition pays first; showing one dissolve — or resist — is a finding either way |
| Discover what your field doesn't do today because it is too expensive — and test whether agents make it possible | The reports never written and the checks never run are invisible until the price drops; this is where genuinely new practice appears |
| Develop frameworks for analyzing a domain's tasks and identifying which are suitable for agents | Every field now faces the mapping problem: how do you take this capacity and figure out where it applies? A framework for that is reusable research |
| Identify gaps in techniques and processes for incorporating agent work at scale in your domain | When the volume of work explodes, the old workflow breaks — someone has to design the new one, the way libraries had to be invented |
| Design methodologies for creating safe toolsets and agent designs for your domain | The tools define what an agent can even attempt; a principled way to design and bound them for your field's risks is a contribution beyond any single agent |
| Design ways to assess and benchmark agents on your domain's work | "Better" means nothing until someone defines it; the field that writes its own measures decides its own standards |
| Study coordination at scale: failures, detection, retry, and orchestration when many agents work in parallel | Scale makes individual mistakes a certainty — this is distributed systems with intelligent actors, and the classic questions (and maybe theorems and bounds) need reinventing |
| Study how errors compound across agent work in your domain — and design the quality control that catches them | When output is cheap, checking becomes the scarce resource; who or what reviews ten thousand results? |
| Design how your domain's knowledge reaches agents: skills, manuals, on-demand training | Fields spent centuries writing for humans; writing the versions an agent reads mid-task is a new genre with its own craft |
| Study the human training component: how people in your field learn to apply agents and the techniques around them | We had to teach everyone to read and write; every domain now needs its equivalent curriculum, and nobody has written it |
| Determine which decisions must stay with a human in your domain — and how agent designs can enforce that | The safety line is domain-specific: your field's ethics, regulation, and stakes decide it, and design has to make it real, not aspirational |
| Analyze the economics: when is agent work worth it in your domain — the cost and quality trades | Tokens, turns, and review burden against hours and salaries; the interesting result is almost always the trade, not the capability |
Notice that most rows are not "build an agent and benchmark it." They are methodologies, frameworks, assessments, and analyses — but every one of them is investigated by building agents and trying things, which is exactly what you will do all week.
After the week
The skeleton you present Friday is the outline of the paper you finish afterward. The completion schedule will be announced by the instructor — think weeks, not months, while the week's momentum is still with you.
And a standing offer: we are happy to keep collaborating — to read drafts, to advise on the study you designed, and to coauthor where the work grows into something shared. The best outcome of this week is not a deliverable; it's a research program that starts here.