Roadmap

1. Chat → Agent
2. MCP & Skills
3. The Forgetting Problem
4. Agent Orchestration
5. The Industry Craze

From Chat to Agent

• A brilliant brain locked in a glass box
• Writes code… but can't run it
• The copy-paste dance: write → error → fix → repeat

Like a consultant who draws blueprints but can't pick up a hammer.

The Agent Arrives

• Same brain, now with hands and feet
• 30 minutes → 30 seconds
• But: confidently generates outdated code from stale training data

The agent needs to be customized for your workflow.

MCP & Skills

MCP — USB standard for AI

• One protocol → plug into WRDS, EDGAR, Stata, Zotero…

Skills — Persistent instructions

• "Winsorize at 1 & 99" — loaded fresh every time

MCP = REACH · Skills = CONSISTENCY

MCP in VS Code

• Each server = a live connection to an external tool
• Community-built — someone builds it, everyone benefits

Practical Uses for Faculty

The Wall: Context Window

• 200K tokens ≈ 200 pages
• Everything goes in — messages, files, code, outputs
• When full → oldest info pushed out

Short-term memory with a hard limit.

The Forgetting Problem

Step 1–3: perfect. Step 4: "What variables? Who are you?"
Not a bug — a fundamental constraint.

One Agent Can't Build a Skyscraper

• One agent = one context window
• Can't juggle every role at once
• Humans solved this long ago: specialize and delegate

Agent Orchestration

• Instead of one super-agent → build a company
• Each specialist gets a fresh context window
• Does one job → returns a summary
• Conductor only sees summaries — clean context

The Research Paper Factory

                         YOU — The CEO
                              │
    ┌───────────┬─────────────┼─────────────┬───────────┐
    │           │             │             │           │
 Strategist  DE-Conductor  DA-Conductor  PW-Editor  Rev-Strategist
 VP Research  VP Data      VP Analysis   VP Writing  VP Revision
    │           │             │             │           │
  Scout       Miner        Builder      Drafter     Respondent
  Analyst     Refiner      Executor     Biblio.     Auditor
              Sentinel     Vizmaker     Typeset     Liaison
              Debugger     Reviewer     Compiler
                                        Submission, Imager

26 agents. You talk to 5 or 6.

The 6 Conductors — Who You Talk To

You give instructions to these 6. They manage everything below.

Shared Services — Defined Once, Used Everywhere

The "SS-" agents live at user level and serve multiple departments:

Real Example — Building a Literature Knowledge Space

This Week: Surfacing Latest Research Trends

• DE-Conductor → scrapes Google Scholar + SSRN via MCP servers
• SS-Analyst → clusters papers by topic, flags emerging themes
• DA-Builder → builds a searchable knowledge graph in DuckDB
• SS-Scribe → generates a structured reading list with summaries

Each agent sees only its own task. Summaries flow up to the conductor — raw code and data never leave the subagent's context.

Key Design Principles

1. Delegation, not conversation

• Task → execute → concise report → clean context

2. Document-based coordination

• Backbone docs (~500 tokens) replace monolithic context scanning — enables parallel execution

3. Multi-model diversity

• Opus for reasoning · Codex for code · Sonnet for review · Gemini for search

Dead Reckoning: Why Oversight Matters

• Compass error of 0.5° → after 3 days, wrong continent
• LLMs generate approximations — chain 50 steps → errors compound
• Real case: agent silently dropped 30% of samples at step 47

Without feedback, small errors become total failure.

The HAHA Method

Human → AI → Human → AI

 ┌────────┐    ┌────────┐    ┌────────┐    ┌────────┐
 │ HUMAN  │───►│   AI   │───►│ HUMAN  │───►│   AI   │───► ...
 │  Set   │    │Execute │    │Validate│    │ Refine │
 └────────┘    └────────┘    └────────┘    └────────┘

• Converts unsupervised → supervised
• Each checkpoint = a labeled correction
• Every major phase → mandatory human review

The feedback loop is everything.

Vertical Integration: The Industry Playbook

What This Means for Us

          ┌─────────────────────────────────┐
          │   Right Questions + HAHA         │  ◄── Human Edge
          ├─────────────────────────────────┤
          │   Scaffolding — Agent Design     │  ◄── Differentiator
          ├─────────────────────────────────┤
          │   Tools — Everyone Has ChatGPT   │  ◄── Table Stakes
          └─────────────────────────────────┘

"Everyone can drive a Ferrari. Not everyone can race F1."

Call to Action

1. Start small — custom instructions for one task
2. Add MCP — connect one external tool
3. Think like a CEO — design delegation

AI doesn't replace researchers. It amplifies them — with the right design.

Thank you · Questions welcome