Agentic Research Project Refinement & Prototyping Session

75 min

Overview

Last session for building. Your team has Paper Skeleton v1, a ranked gap list topped by three reviewer questions, and a morning's worth of architecture — planning, context boundaries, coordination patterns, multiple agents. This session does three things, in strict priority order: it forces the architecture question on your own project ("do we actually need any of this?"), it closes the top gap with one targeted experiment, and it turns the skeleton into v2 plus the presentation you give at 14:00. Budget hard: twenty minutes, thirty, twenty-five. The presentation is the deliverable; polish on anything else is time stolen from it.

The honesty rule

"Do we need an agent at all?" applies to your own project most of all. A paper that shows a single well-prompted loop suffices where everyone assumed a multi-agent system is a STRONGER result, not a weaker one. Add planning, memory, or a second agent only when you can say which failure from your own trials it prevents.

The session at a glance

1
Architecture check: walk your approach through the week's decision points.
2
If coordination is warranted — prototype the conversational structure, cheaply.
3
Close the top gap: one targeted Forge run against your strongest reviewer question.
4
Update skeleton to v2 and write Future Work.
5
Build the presentation from the fixed slide template.

Part 1 · The architecture check — twenty minutes

Walk your approach through the forks this week taught, and write one line per verdict — these lines go straight into the paper's Approach section as design rationale:

The decision points

agent at allcould a single well-designed prompt, or a plain script, do what your approach needs? What in your notebook says otherwise?

planningdoes your task need a plan surfaced before acting — and did your trials show reactive trajectories failing in ways a plan would catch?

context boundarieswhat must be kept OUT of the main conversation? Which of your two loads — task and coordination — was actually heavy in your runs?

more than one agentname the pattern if you claim one — explorer, scientist/designer, planner–executor, actor–critic — and name the failure it prevents. "It seems more powerful" is not a failure it prevents

safetywhich decisions must stay with a human, and does the architecture enforce that or just intend it?

If a second agent survives that test, spend ten minutes prototyping the conversational structure in the Forge or on paper: who speaks first, what crosses the boundary between the agents, what each one must never see. The message contents are the design — two exchanged turns on paper beat a diagram.

Part 2 · Close the top gap — thirty minutes

Take the top item on the gap list — usually the strongest reviewer question from yesterday — and chase it with one more quick build or trial in the Forge: one case, one comparison, tallied the same rough, honest way as yesterday, excerpt captured. Add the row to your proof-of-concept tally and the line to your notebook.

One gap, closed, beats three gaps wounded

Resist splitting the time across questions. Friday's audience will forgive open gaps — the gap list is future work. What they remember is the question you answered with something that RAN.

Part 3 · Skeleton v2 and the presentation — twenty-five minutes

Fold everything in: the architecture verdicts into Approach, the new row into Preliminary Experiments, the remaining gap list — reframed positively — into Future Work. That is Skeleton v2. Then build the presentation:

The presentation — seven slides, eight minutes

1 · the capacitywhat becomes cheap in your domain, and where value migrates — thirty seconds, this is your hook

2 · the problem & the evidencethe chosen problem and your strongest numbers from the Positioning Memo

3 · how it's done todaycurrent practice, its cost, and the borrow/reinvent/invent verdict — with the invent column starred

4 · the approachyour action language and architecture — with the one-line rationale for every mechanism you kept OR rejected

5 · how you'll knowmeasures, baselines, the mini-benchmark — one slide, the methodology at a glance

6 · what the proof of concept showedthe tally and your best trajectory excerpt — real data, modest claims

7 · what's nextfuture work from the gap list, and the first experiment you'd run after this week

Before the presentations

💾 Capture

Ready to go at 14:00:

Write down:

1The seven slides.
2Skeleton v2 — sent to your designated discussant team so they can read it before you present.
3One line: the mechanism you REJECTED in Part 1, and why — you will be asked.

Be ready to discuss:

While building slide 4, check yourself once more against the honesty rule: every box in your architecture should be traceable to a failure it prevents. Your discussants have the same checklist — because you wrote it with them all week.