IDE Agent Collaboration Through a Case Study
20 February 2026 (Thu) · 10:00 AM HKT · Online via Zoom
Use ↓ or scroll to navigate · Press T to edit & download your own copy
We demo Agentic AI in an IDE using a real project about math research storytelling — the KissingNumber case study. You watch a 17-step process log showing every interaction between a human and an AI agent, from reading an SCMP article to drafting a letter to the editor.
After the demo, you complete a short exit questionnaire (~3 min). Students who need CCL credits will receive them after completing the questionnaire.
Then we start the hands-on session — you try the tools yourself. We can take questions while you work.
IDE agents are not chatbots. They read your files, run your code, and build alongside you. Think of them as a tireless junior colleague who never sleeps.
Both AI agents and chatbots are powered by large language models — but there is a fundamental difference:
Can only respond with text (including images & videos). You copy-paste content in and out. One task at a time. Limited context window.
Can invoke tools, read & edit files, run code, search the internet, and perform multi-step tasks directly on your project folder.
An IDE (Integrated Development Environment) is a tool traditionally used by programmers. But our workshop is not about programming — it is about using AI agents inside IDEs to handle everyday document tasks:
VS Code (Microsoft, free & open-source) · Cursor (built on VS Code) · Trae (free AI IDE, built on VS Code) · Replit (cloud-based, no install needed)
IDEs are not just for programmers. If you can write a text message, you can use an AI agent in an IDE. The innovation is breaking out of fixed mental frameworks — just like discovering that 5 − 1÷5 × 5 = 24.
Most people know 年 as "year." But 年 was originally a monster — 年兽 (Nian Shou). It comes when we least expect it.
AI is our 年兽. While we discuss whether to adopt it, it is already transforming the world.
Use a compelling metaphor or story to motivate action. The 年兽 framing turns abstract urgency into a vivid narrative your audience remembers.
On 19 February 2026 (yesterday), the SCMP reported on a breakthrough by researchers from Peking University, Fudan University, and the Shanghai Academy of AI for Science. Their AI system PackingStar used game-theoretic reinforcement learning to find record-breaking solutions to the kissing number problem — a 300-year-old geometry question: how many identical spheres can touch a central sphere in n dimensions?
The team described their work as a “romance” between scientists and AI: the AI discovered configurations beyond human geometric intuition, while the mathematicians provided verification and further improvements.
Our goal: Use an AI agent in an IDE to understand this research and write a letter to the SCMP editor in response. The first step: the SCMP article mentions the research but does not link to the paper, so we ask the agent to locate it on arXiv.
given the report, search arXiv to locate the paper
published in Nov 2025 (around the time)
download the html or latex version and place the
paper in a sub-folder /paper
create process.log and update in real time what is doneThe agent used the arXiv API to search, locate arXiv:2511.13391, and download the full paper — no manual browsing needed. 17 steps were logged in real time.
Agent as researcher: When a news article doesn’t link to the original source, give the article to your AI agent and ask it to locate the paper. The agent can query APIs (arXiv, PubMed, Google Scholar) directly. Keep a real-time process log to make your work reproducible.
The SCMP report was already in the project as a reference file. After searching arXiv, the agent downloaded the full paper into a local sub-folder. Now the project has both sources. This is a key advantage of an AI agent over a chatbot:
Here is the actual prompt (Brainstorm.md) given to the agent — notice how it references multiple local files and asks for three outputs at once:
I am Simon Wang the author of the SCMP letter on solar weather
genAI2026/.../Ref/solarWeatherLetter.md
I intend to write another SCMP letter in response to
genAI2026/.../Ref/SCMPreport.md which is about this study
reported in the paper here genAI2026/.../paper
I am brainstorming some points to make in the letter for general
public; I do not have training on advanced math; so ideas have
to be accessible to non math experts
also check out how I got help from AI to write a different letter
here https://poe.com/s/JdBL1zQr2xd9jamabP16 and the published
letter https://www.scmp.com/opinion/letters/article/3342445/...
create a partnerAIwriting.md to describe the process and identify
some transferable lessons
read through all the docs and generate a ideas4letters.md in bullet
points to discuss some potential insights to share in the letter
(try to identify some insights that can be found from analyzing the
preprint that have not been covered in the SCMP report)Note: solarWeatherLetter.md referenced above is a copy of this published SCMP letter (2019) — included in the project folder so the agent can study the author’s previous writing style.
Multi-task prompting: Unlike a chatbot, an IDE agent can read all your local files and execute multiple tasks in one go. Give it references to your existing files and ask for several outputs — the agent maintains context across all of them.
The agent created partnerAIwriting.md by analyzing the Poe conversation and the published SCMP letter, then documented a reusable 5-step process:
Transferable lessons:
- Use a single, shareable thread
Keeping one Poe (or similar) conversation for a letter makes
it easy to revisit, refine, and later document "how this was
written."
- Lead with the hook and your view
Start by pasting the article (or summary) and 2-3 sentences
on what you want to say. That keeps the AI from drifting.
- Constrain length and structure early
Tell the AI the word limit and preferred shape. Reduces
back-and-forth later.
- One main ask
The tax/graduates letter centres on one policy idea. One clear
ask is easier for readers and editors.
- Reuse the same workflow for new letters
Same steps: anchor in the report, state your angle, draft with
AI, then fact-check and edit.
- Acknowledge AI assistance if you wish
Depending on outlet policy, you can mention the letter was
drafted with AI assistance.Document your process: Ask the agent to analyze how you worked with AI before and create a reusable process document. This turns ad-hoc collaboration into a repeatable method you can share with others.
The agent created ideas4letters.md by reading all source files. It found insights from the preprint not covered in the SCMP report:
Insights from the preprint NOT fully in the SCMP report:
- Two agents, one goal: PackingStar is a two-player game;
one agent fills a matrix, the other corrects mistakes.
"Like two experts refining a shared blueprint."
- AI broke old assumptions: New configurations "challenge
long-held antipodal paradigms" — AI found patterns that
don't follow the symmetry rules mathematicians assumed.
- Humans then did better: "Inspired by these patterns,
humans devised further improved constructions" — a genuine
feedback loop, not replacement.
- Proof still missing: For 25D, "a rigorous mathematical
proof is still lacking." AI expands the frontier;
mathematicians still provide certainty.
- Scale of discovery: "Over 6000 new structures"
independently validated and included in standard databases.
Possible angles for a letter (pick one or combine):
1. Collaboration, not replacement
2. Why Hong Kong / the region should care
3. Trust and verification
4. From 1694 to 2025
5. One concrete "so what"What documents do you have that an AI agent could cross-reference? A news article + a research paper + your own notes = the agent finds connections you might miss. How would you apply this input → process → output model to your own problem?
The PackingStar team published a 16-minute video on the PKU WeChat channel where the lead researcher explains the project in Chinese. The agent processed this video to extract insights not found in the paper or SCMP report:
process_session.py pipeline: keyframe extraction (20s intervals + hash dedup) → OCR (EasyOCR) → transcript parsing & alignment → notes generationmoreIdeas.md from video transcript — captured unique insights only available in the video: digital life, 14D nightmare, 无中生有 (creating from nothing), the origin of the two-agent design from "manually deleting bad balls"Video-only angles not found in paper or SCMP report:
- "创造新的数字生命" (creating new digital life)
The lead researcher frames PackingStar as creating
something from nothing (无中生有) — unlike AlphaFold
which uses existing data, this AI creates new structures.
- "14D nightmare" (14维的梦魇)
To validate the system, the team insisted on reproducing
the known 14D result (1932 kissing number). It took
enormous effort — described as a nightmare — but proved
the system could match human mathematics.
- Origin of two-agent design
A mathematician found he needed to "manually delete bad
balls" (手动删掉坏球) for the system to work — one bad
ball ruins everything (一粒老鼠屎). This human act was
encoded as the second agent (Corrector).
- "解释神" (interpreter of the gods)
The lead researcher says the work he spent the most time
on was "interpreting the data from the AI black box."
- "射箭到月亮" / AlphaGo "神之一手"
The AI's discoveries are compared to "shooting an arrow
to the moon" — configurations beyond human intuition.Pipeline command:
python process_session.py <Ref_path> PKU_wechat.mp4 PKU_wechat.txtOutputs: keyframes/, keyframes.json, ocr_results.json, aligned_data.json, ChineseNotesCoze.md
Multimedia content analysis: Use existing scripts (or ask the agent to write one) to extract keyframes and OCR text from videos. Then ask the agent to cross-reference the transcript with written sources to find unique insights. This works for lecture recordings, conference talks, interviews, etc.
Three rounds of HTML presentations, each driven by specific human feedback:
View each version as a presentation:
View explain.html (V1) ↗The original prompts that drove each iteration:
V1: "create an HTML explainer from scratch; describe novelty, link to SCMP report and PKU video; letter focus on human value"
V2: "more background on the paper; more on PackingStar; label three sources of human value (historical, problem, game-AI); look at GitHub code; add letter outline up to 150 words"
V3: "standalone, consolidate previous two versions, accessible to non-math people; can AI prove the result? check for other AI models that can help prove"
explain.md (original prompt) ↗Iterative content creation with specific feedback: Don't aim for perfection in one go. Each iteration should include specific instructions: "more background on X," "make it accessible to non-experts," "add section on Y." Vague feedback produces vague improvements. The agent maintains consistency across versions while responding to your precise direction.
Three drafts, each driven by human critical judgment:
The user asked the agent to provide exact sources for every claim. The agent traced each sentence to paper sections and PKU video timestamps:
| "overturned assumptions" | Paper §Intro ¶4: "fundamentally challenge long-held antipodal paradigms" |
| "interpreting the AI black box" | PKU video ~15:22: 解释神 — "the work I spent the most time on" |
| "second agent from human instinct" | PKU video ~09:35 + Paper §Intro ¶3 |
| "humans devised further improved constructions" | Paper Abstract + §Intro ¶4 |
I refer to your report on the PackingStar AI system
and its advances on the kissing number problem
(February 9). Your headline calls this a "romance,"
and the metaphor runs deeper than it might seem.
PackingStar discovered sphere arrangements that
overturned assumptions mathematicians had held for
decades. But the system itself was born from human
instinct: a mathematician found he needed to manually
delete bad entries for the AI to make progress, so the
team taught their machine to mimic that act — encoding
a human correction into the AI's architecture.
But the machine returned the favour. Its configurations
"fundamentally challenge long-held antipodal paradigms,"
forcing mathematicians to rethink what they believed
about high-dimensional symmetry. Inspired by those
surprises, humans then extracted new construction
principles and pushed results further still.
That is what makes this a genuine romance, not mere
human oversight of a fast calculator: each partner
transforms the other. As AI reshapes research, we
should invest not only in faster algorithms but in the
human expertise to question, interpret, and build on
what AI reveals.The revision prompt that started the drafting process:
revise.md (drafting prompt) ↗Critical feedback drives quality: When a draft feels generic, tell the agent why ("readers will say 'of course'"). The agent can re-read all sources for non-obvious evidence.
Source verification: Ask the agent to cite the exact paper section, page, or video timestamp for every claim. This turns vague assertions into well-sourced arguments.
Your own insight matters: The "romance" rewrite came from the human noticing that the SCMP headline was structurally accurate. The agent executed the rewrite, but the creative leap was human.
The PackingStar team open-sourced their code and data on GitHub. The agent cloned this repository and performed a deep analysis:
ReadmeSimon.md with data file table, two representations, verification scripts, and 8 visualization ideasVerification scripts found in repo:
verify_coordinates.py # Check coordinate validity
verify_cosmatrix.py # Check cosine matrix properties8 visualization ideas generated:
Codebase exploration: Ask the agent to clone a repository, explore its structure, inspect data files, and summarize what it finds. It can generate documentation, identify patterns, and propose visualization strategies — turning opaque code into understandable insights.
The agent generated all 8 planned visualizations, explained existing diagrams, and assembled a final 23-slide visual narrative:
| 1. Cosine distribution | 12D has only 3 distinct cosine values — hidden algebraic regularity |
| 2. Gram matrix heatmap | Block-diagonal patterns reveal internal sub-structures |
| 3. Eigenvalue spectrum | Rank = 12, 13 confirms dimensionality |
| 4. PCA 3D projection | 17D spheres show layered shells in projection |
| 5. Contact graph | 12D: every sphere has degree 36 (strongly regular) |
| 6. Old vs new chart | ΔK grows from +8 (25D) to +5,476 (31D) |
| 7. Partitioned sub-structures | Triangle-assembly pattern for 29D/31D records |
| 8. Rational 13D | 6 exact rational cosine values: {−1, −1/2, −1/4, 0, 1/4, 1/2} |
A standalone 23-slide visual story using PKU video keyframes and all 8 visualizations. Includes the final letter draft and a "How AI Helped Me Write This" transparency slide showing 19 human/AI steps.
Narrative arc: question → cosine idea → single player → 14D nightmare → two players → fragments → results → visualizations → creative loop → letter → meta-lesson
View explain04.html ↗ | GitHub Pages (live) ↗Data visualization from code: The agent can read NumPy data files, generate Python scripts, run them, and explain the output. This transforms raw research data into presentation-ready figures.
Visual storytelling: Once you have all assets (keyframes, visualizations, text), ask the agent to assemble a narrative presentation. It can weave multiple media into a coherent story with navigation.
Please complete the exit questionnaire now.
Your responses contribute to CCL credit hours for this workshop.
Takes approximately 3–5 minutes to complete
Just as PackingStar's breakthrough came from a creative loop between AI and mathematicians, this letter itself was produced through the same kind of loop: the human provided the angle, the critical pushback, and the "romance" insight — the agent provided research at scale, source verification, and multiple drafts. Each partner transformed the other's thinking.
Before the hands-on session, set up one of these free tools on your computer:
Sign up for GitHub using your institutional email (e.g. your @hkbu.edu.hk address), then apply for education benefits:
Left panel — File explorer (view & manage folders) Centre — File editor (top) & Terminal (bottom) Right panel — AI agent chat
As a workshop participant, you can claim a 3 days of free Trae Pro tokens. Enter your email below to receive your unique code.
Session 2 of 7 · 5 March 2026 (Thursday) · 7:30 – 9:30 PM · Online via Zoom
Learn how to use AI agents to turn published academic papers into textual mentors that guide your own research and writing.
Open to HKBU and other university students and scholars