MCP DEV SUMMIT SEOUL · OFFICIAL WORKSHOP

How to Use
Sol-Terra-Luna
in One Goal

3 × 25 MINUTES · GOAL · ROUTING · UNDERSTANDING

정구봉 · GOOBONG JEONGTEAM ATTENTION 창업자 · CODEX AMBASSADOR
GOOBONG JEONGTEAM ATTENTIONCODEX AMBASSADOR3 × 25 MIN
MCP DEV SUMMIT SEOUL · OFFICIAL WORKSHOP

How to Use
Sol-Terra-Luna
in One Goal

3 × 25 MINUTES · GOAL · ROUTING · UNDERSTANDING

GOOBONG JEONG · 정구봉FOUNDER, TEAM ATTENTION · CODEX AMBASSADOR
GOOBONG JEONGTEAM ATTENTIONCODEX AMBASSADOR3 × 25 MIN
TODAY · 3 × 25 MINUTES

세 번의 세션,
하나의 운영 시스템

Goal로 일을 정의하고, 모델에 라우팅하고, 인간의 이해를 깊게 만드는 75분입니다.

01 · 25 MINGoal을 운영 단위로

프롬프트가 아니라 결과·증거·제약을 유지하는 루프를 설계합니다.

검증 가능한 /goal 1개
02 · 25 MINSol · Terra · Luna 라우팅

실패 비용과 검증 가능성으로 판단·구현·반복을 배분합니다.

실제 작업의 라우팅 맵
03 · 25 MIN이해를 위한 Harness

팟캐스트·Companion·시뮬레이션으로 실행 이후의 병목을 다룹니다.

개인 학습 루프 1개

각 세션은 설명 → 운영 패턴 → 바로 적용할 산출물로 끝납니다.

TODAY · 3 × 25 MINUTES

Three sessions.
One operating system.

In 75 minutes, we will define work with Goals, route it across models, and deepen human understanding.

01 · 25 MINGoal as the operating unit

Design a loop that preserves outcome, evidence, and constraints—not just a prompt.

ONE VERIFIABLE /GOAL
02 · 25 MINRoute Sol · Terra · Luna

Allocate judgment, implementation, and repetition by failure cost and verifiability.

A ROUTING MAP FOR REAL WORK
03 · 25 MINA harness for understanding

Use podcasts, companions, and simulations to attack the bottleneck after execution.

ONE PERSONAL LEARNING LOOP

Every session ends with explanation → operating pattern → an artifact you can use immediately.

SESSION 01 · 25-MINUTE MAP

Goal을 잘 돌리는
운영 계약

25분 뒤, 모호한 작업 하나를 증거로 종료할 수 있는 /goal로 바꿀 수 있습니다.

25MINUTES
ONE OUTCOME
01 · WHY GOALPrompt에서 Loop로

코딩 에이전트가 지속되는 Goal로 수렴하는 이유

02 · WRITE THE CONTRACT완료를 먼저 정의

최종 상태 · 증거 · 제약 · 중단 규칙

03 · RUN THE LOOP실행하고 증명하기

Execute → Score → Check → Route

TAKEAWAY · /goal [완료 상태] — [증거]로 검증 — [제약]을 보존

SESSION 01 · 25-MINUTE MAP

Run Goals with
an operating contract

After 25 minutes, you can turn one vague task into a /goal that ends on evidence.

25MINUTES
ONE OUTCOME
01 · WHY GOALFrom prompt to loop

Why coding agents are converging on durable Goals

02 · WRITE THE CONTRACTDefine done first

End state · evidence · constraints · stop rules

03 · RUN THE LOOPExecute and prove

Execute → Score → Check → Route

TAKEAWAY · /goal [end state] — verified by [evidence] — preserving [constraints]

00 · TEAM ATTENTION

AI Native Builders
→ Founders.

각자의 회사를 만들고, 서로의 판단을 검토합니다.

  1. 01NAVER D2SF 강남에서 함께 일하기
  2. 02모든 founder가 자기만의 story 만들기
  3. 03BUILD · WORK OUT · CHECK IN
RESIDENCY · ROUTINE
RESIDENCY · ROUTINE
FOUNDERS · HIR
FOUNDERS · HIR
REAL-WORLD GATHERINGS
REAL-WORLD GATHERINGS
00 · TEAM ATTENTION

AI Native Builders
→ Founders.

Build your company. Review decisions together.

  1. 01Work together at NAVER D2SF Gangnam
  2. 02Every founder builds a distinct story
  3. 03BUILD · WORK OUT · CHECK IN
RESIDENCY · ROUTINE
RESIDENCY · ROUTINE
FOUNDERS · HIR
FOUNDERS · HIR
REAL-WORLD GATHERINGS
REAL-WORLD GATHERINGS
01 · FOUNDER COMMUNITY AS A LAB

도시와 도메인을
바꿔가며 실험했습니다.

SF · Singapore · Seoul · Dosan
해커톤에서 논문 작성까지

SEOUL
SEOUL
SAN FRANCISCO
SAN FRANCISCO
SINGAPORE
SINGAPORE
01 · FOUNDER COMMUNITY AS A LAB

We changed cities—
and changed domains.

SF · Singapore · Seoul · Dosan
From hackathons to writing papers

SEOUL
SEOUL
SAN FRANCISCO
SAN FRANCISCO
SINGAPORE
SINGAPORE
02 · WE BUILD PAST THE PROMPT

에이전트를 쓰는 데서
멈추지 않습니다.

Codex와 함께 Model Finetune · GAME WORLD MODEL

김승진 · BLIT · TEAM ATTENTION 2026 S2
15H / ROUND× 20 ROUNDS≈ 300H A100
BLIT · ASSET GENERATION
BLIT · ASSET GENERATION
MODEL EVALUATION LAB
MODEL EVALUATION LAB
MASK + SPILL RESTORATION
MASK + SPILL RESTORATION
02 · WE BUILD PAST THE PROMPT

We build
past the prompt.

Model finetuning with Codex · GAME WORLD MODEL

Seungjin Kim · BLIT · TEAM ATTENTION 2026 S2
15H / ROUND× 20 ROUNDS≈ 300H A100
BLIT · ASSET GENERATION
BLIT · ASSET GENERATION
MODEL EVALUATION LAB
MODEL EVALUATION LAB
MASK + SPILL RESTORATION
MASK + SPILL RESTORATION
SESSION 01 · CODING AGENT CONVERGENCE

프롬프트 엔지니어링에서
루프 엔지니어링으로

설계 단위가 한 번의 지시에서, 실행하고 증거를 확인하며 다음 행동을 결정하는 시스템으로 이동합니다.

01PROMPT ENGINEERING

한 번의 지시를 최적화

산출물 · 더 나은 첫 시도
02GOAL CONTRACT

결과·증거·제약을 정의

산출물 · 검증 가능한 완료 정의
03LOOP ENGINEERING

실행 → 평가 → 계속·중단·승격

산출물 · 시간에 따라 신뢰할 수 있는 진전

오늘의 실습 · 내 작업 하나를 검증 가능한 /goal로 바꿉니다.

SESSION 01 · CODING AGENT CONVERGENCE

From prompt engineering
to loop engineering

The design unit shifts from one instruction to a system that executes, checks evidence, and decides what happens next.

01PROMPT ENGINEERING

Optimize one instruction

OUTPUT · A better first attempt
02GOAL CONTRACT

Define outcome, evidence, and constraints

OUTPUT · A testable definition of done
03LOOP ENGINEERING

Execute → evaluate → continue, stop, or escalate

OUTPUT · Reliable progress over time

TODAY’S WORKSHOP · Turn one real task into a verifiable /goal.

03 · THE INTERFACE IS CONVERGING

서로 다른 도구가
지속되는 Goal로 수렴합니다.

Goal은 여러 단계와 세션에 걸쳐 목표, 완료 기준, 증거, 중단·에스컬레이션 규칙을 유지합니다.

CODEX

제품 안에 지속되는 objective

CLAUDE CODE

Stop hook이 지키는 세션

CUSTOM HARNESS

계획·실행·증거 workflow

→→→
01OBJECTIVE02DEFINITION OF DONE03REQUIRED EVIDENCE04STOP OR ESCALATE

도구는 다릅니다. 반복해서 나타나는 운영 계약은 같습니다.

질문 · 지금 당신의 작업은 무엇이 되면 정말 끝나는가?

03 · THE INTERFACE IS CONVERGING

Different tools converge on
the same durable Goal.

A Goal carries the objective, completion test, evidence, and stop or escalation rules across steps and sessions.

CODEX

Persistent objective in the product

CLAUDE CODE

Session guarded by a stop hook

CUSTOM HARNESS

Plan–execute–evidence workflow

→→→
01OBJECTIVE02DEFINITION OF DONE03REQUIRED EVIDENCE04STOP OR ESCALATE

THE TOOLS DIFFER. THE RECURRING OPERATING CONTRACT IS THE SAME.

PROMPT · What would make your task truly done?

04 · GOAL CONTRACT

좋은 Goal은
‘계속해’가 아닙니다.

완료를 선언할 수 있는 조건을 먼저 씁니다.

01DESIRED END STATE02SPECIFIC EVIDENCE03CONSTRAINTS · BOUNDARIES04NEXT-ACTION RULE05BLOCKED REPORT
/goal [최종 상태]를 완성하라.
[증거]로 검증하고 [제약]을 보존하라.
매 반복에서 다음 행동을 선택하라.
유효한 경로가 없으면 blocker와 uncertainty를 분리해 보고하라.
04 · GOAL CONTRACT

A good Goal is not
“keep going.”

Write the conditions that justify declaring completion.

01DESIRED END STATE02SPECIFIC EVIDENCE03CONSTRAINTS · BOUNDARIES04NEXT-ACTION RULE05BLOCKED REPORT
/goal Complete [end state].
Verify with [evidence] while preserving [constraints].
Choose the next action after every iteration.
If no valid path remains, report blockers and uncertainty separately.
05 · THE RUNTIME

Goal 런타임은 증거가 통과하거나
유효한 경로가 사라질 때까지 반복합니다.

각 반복은 행동이 아니라 검증 가능한 상태 전이를 만듭니다.

01EXECUTE

다음 bounded action 실행

artifact 또는 change
→
02SCORE

결과를 성공 기준과 비교

score와 gap
→
03CHECK

증거와 제약을 검증

pass · fail · blocked
→
04ROUTE

통과 → 종료 · gap → 계속 · 막힘 → 승격

다음 상태 선택
GOAL · checkout p95 < 120ms→EVIDENCE · benchmark = 146ms→NEXT · CONTINUE · profile DB calls
05 · THE RUNTIME

A Goal runtime repeats until
it passes—or must escalate.

Every iteration produces a verifiable state transition—not just more activity.

01EXECUTE

Perform the next bounded action

Artifact or change
→
02SCORE

Compare result with success criteria

Score and gaps
→
03CHECK

Validate evidence and constraints

Pass · fail · blocked
→
04ROUTE

Pass → terminate · gap → continue · blocked → escalate

Choose next state
GOAL · checkout p95 < 120ms→EVIDENCE · benchmark = 146ms→NEXT · CONTINUE · profile DB calls
06 · CODEX / CLAUDE CODE

같은 계약,
다른 런타임 표면

Codex는 제품 안에서 루프를 관리하고, Claude Code에서는 사용자가 세션·Hook·외부 상태로 루프를 조립합니다.

CODEX GOALCLAUDE CODE HARNESS

제품 안의 Goal 상태

세션 또는 외부 상태

Goal 상태 + 중단 조건

Hook + evaluator / stop script

제품 runtime

사용자 또는 팀

공통 계약 · objective + evidence + stop condition + blocker report

질문 · 당신의 루프는 제품, hook, 외부 harness 중 어디에 사는가?

06 · CODEX / CLAUDE CODE

Same contract,
different runtime surfaces

Codex manages the loop in-product; Claude Code lets users assemble it from sessions, hooks, and external state.

CODEX GOALCLAUDE CODE HARNESS

Goal state in the product

Session or external state

Goal state + stop condition

Hooks + evaluator or stop script

Product runtime

User or team

SHARED CONTRACT · objective + evidence + stop condition + blocker report

PROMPT · Where does your loop live: product, hook, or external harness?

07 · WHAT USERS ACTUALLY WANT

더 긴 대화가 아니라
검증 가능한 완료

증거가 Goal 계약을 충족할 때만 작업이 끝납니다.

GOAL CONTRACT

최종 상태 · 증거 · 제약

→
AGENT LOOP

행동 · 관찰 · 수정

→
EVIDENCE CHECK

증거가 계약을 충족하는가?

YESDONE · 검증된 산출물
NOT YETNEXT ACTION → AGENT LOOP
NO VALID PATHBLOCKER + UNCERTAINTY 보고

질문 · 어떤 증거가 루프를 끝내고, 어떤 증거가 다시 돌려보내는가?

07 · WHAT USERS ACTUALLY WANT

Not a longer chat.
Verifiable completion.

A task is done only when evidence satisfies the Goal contract.

GOAL CONTRACT

End state · evidence · constraints

→
AGENT LOOP

Act · inspect · revise

→
EVIDENCE CHECK

Does evidence satisfy the contract?

YESDONE · verified artifact
NOT YETNEXT ACTION → AGENT LOOP
NO VALID PATHREPORT BLOCKER + UNCERTAINTY

PROMPT · What evidence ends the loop—and what evidence sends it back?

08 · GOALS, NOT RECIPES

범용 앱 위에
명시적인 Harness가 생깁니다.

계획·실행·증거를 제품 밖에서도 강제합니다.

LazyCodexPLAN → WORK → VERIFY
Gajae-CodeSHAPE → ACT → PROVE
08 · GOALS, NOT RECIPES

Explicit harnesses emerge
on top of general agents.

They enforce planning, execution, and proof outside the core product.

LazyCodexPLAN → WORK → VERIFY
Gajae-CodeSHAPE → ACT → PROVE
09 · THREE CONNECTED LAYERS

운영 패턴은
서로 연결된 세 층입니다.

각 층은 위의 정의를 실행하고 검증할 수 있는 루프로 바꿉니다.

01완료를 정의GOAL CONTRACT

완료 상태 · 증거 · 제약

/goal
02증거를 생산HARNESS WORKFLOW

계획 · 실행 · artifact · 검증

ulw / gjc
03다음 상태를 선택LOOP ENGINEERING

계속 · 종료 · 사람에게 승격

review / QA
완료가 모호하다 → 01실행이 흔들린다 → 02품질이 무너진다 → 03

워크숍 · 지금 당신을 막고 있는 층 하나에 표시하세요.

09 · THREE CONNECTED LAYERS

The operating pattern has
three connected layers.

Each layer turns the one above it into an executable, verifiable loop.

01DEFINES DONEGOAL CONTRACT

End state · evidence · constraints

/goal
02PRODUCES PROOFHARNESS WORKFLOW

Plan · execute · artifacts · checks

ulw / gjc
03CHOOSES NEXT STATELOOP ENGINEERING

Continue · terminate · escalate to a human

review / QA
DONE IS VAGUE → 01EXECUTION DRIFTS → 02QUALITY DEGRADES → 03

WORKSHOP · Mark the one layer blocking you now.

10 · OPERATING PRINCIPLE

/goal은 명령어가 아니라
검증 가능한 끝까지 가는 운영 단위입니다.

결과를 정하고 · 증거로 묶고 · 루프를 계속 설계합니다.

01GOAL CONTRACT

LLM과 명시적인 완료 계약을 맺습니다.

END STATE · EVIDENCE · CONSTRAINTS
02OPERATING LOOP

계획·실행·증거를 도구 밖에도 명시합니다.

PLAN · WORK · PROOF
03HUMAN ENGINEER

예외를 검토하고 현실이 바뀌면 계약을 고칩니다.

REVIEW · QA · STOP

지금 작성 · /goal [완료 상태] — [증거]로 검증 — [제약]을 보존

10 · OPERATING PRINCIPLE

/goal is not a command.
It carries work to a verified end.

Set the outcome. Bind it to evidence. Keep engineering the loop.

01GOAL CONTRACT

Make an explicit completion contract with the LLM.

END STATE · EVIDENCE · CONSTRAINTS
02OPERATING LOOP

Make plan, execution, and evidence explicit outside the tool.

PLAN · WORK · PROOF
03HUMAN ENGINEER

Review exceptions and revise the contract when reality changes.

REVIEW · QA · STOP

WRITE NOW · /goal [end state] — verified by [evidence] — preserving [constraints]

SESSION 02 · SOL / TERRA / LUNA

한 모델을 고르는 대신
일을 라우팅합니다.

SOL ORCHESTRATES · TERRA BUILDS · LUNA MULTIPLIES

CLUSTERROUTEPROVEESCALATE
SESSION 02 · SOL / TERRA / LUNA

Stop choosing one model.
Start routing work.

SOL ORCHESTRATES · TERRA BUILDS · LUNA MULTIPLIES

CLUSTERROUTEPROVEESCALATE
SESSION 02 · 25-MINUTE MAP

세 모델을
하나의 Goal로 운영하기

25분 뒤, 실제 작업 하나를 Sol·Terra·Luna에 나누고 다시 통합하는 라우팅 맵을 만들 수 있습니다.

25MINUTES
ONE OUTCOME
01 · ASSIGN ROLES서열이 아닌 역할

Sol의 판단 · Terra의 구현 · Luna의 반복

02 · ROUTE THE WORK두 축으로 선택

실패 비용 × 자동 검증 가능성

03 · ESCAPE ONE LOOPCommand Center로 확장

한 세션에서 여러 도메인 세션으로

TAKEAWAY · 한 실제 요청을 Cluster → Model → Evidence → Escalation으로 그리기

SESSION 02 · 25-MINUTE MAP

Operate three models
inside one Goal

After 25 minutes, you can split one real task across Sol, Terra, and Luna—and integrate it again.

25MINUTES
ONE OUTCOME
01 · ASSIGN ROLESRoles, not rankings

Sol for judgment · Terra for building · Luna for iteration

02 · ROUTE THE WORKChoose on two axes

Failure cost × automatic verifiability

03 · ESCAPE ONE LOOPExpand to a command center

From one session to multiple domain sessions

TAKEAWAY · Map one real request as Cluster → Model → Evidence → Escalation

11 · THREE DURABLE TIERS

세 모델은 서열이 아니라
서로 다른 운영 역할입니다.

SOL

판단 · 설계 · 통합

HIGHEST CAPABILITY
TERRA

구현 · 분석 · 검증

$2 / $12
LUNA

탐색 · 변환 · 반복

$0.20 / $1.20
OPENAI · API PRICE PER 1M INPUT / OUTPUT TOKENS · 2026-07-30
11 · THREE DURABLE TIERS

Three models are not a ranking.
They are operating roles.

SOL

JUDGMENT · ARCHITECTURE · SYNTHESIS

HIGHEST CAPABILITY
TERRA

IMPLEMENTATION · ANALYSIS · VERIFICATION

$2 / $12
LUNA

SEARCH · TRANSFORM · REPEAT

$0.20 / $1.20
OPENAI · API PRICE PER 1M INPUT / OUTPUT TOKENS · 2026-07-30
12 · LUNA = ITERATION VOLUME

가격이 80% 내려가도
일의 중요도가 80% 줄지는 않습니다.

탐색·포맷 변환·대량 검증을 더 자주 반복할 수 있습니다.

80%
LUNA API 가격 인하
OPENAI · ADVANCING THE PRICE-PERFORMANCE FRONTIER WITH GPT-5.6 · 2026-07-30
12 · LUNA = ITERATION VOLUME

An 80% price cut does not make
the work 80% less important.

It buys more search, transformation, and large-batch verification.

80%
LUNA API PRICE REDUCTION
OPENAI · ADVANCING THE PRICE-PERFORMANCE FRONTIER WITH GPT-5.6 · 2026-07-30
13 · ONE REQUEST, FOUR WORKSTREAMS

‘발표 자료 만들어줘’는
한 작업이 아닙니다.

서사와 기준

무엇을 남길지 결정

원본 탐색

파일·웹·대본 수집

구현과 테스트

HTML·내비·렌더

최종 Taste

버릴 것과 고칠 것 판단

13 · ONE REQUEST, FOUR WORKSTREAMS

“Build the deck”
is not one task.

Narrative + bar

Decide what matters

Source mining

Collect files, web, scripts

Build + test

HTML, navigation, renders

Final taste

Judge what to cut and fix

14 · SOL’S FIRST JOB

Sol의 첫 임무는 코딩이 아니라
작업의 모양을 결정하는 것입니다.

01CLUSTER
독립 작업 찾기
→
02RISK
실패 비용 붙이기
→
03MODEL
실행 tier 선택
→
04EVIDENCE
검증 표면 정의

난이도가 아니라 실패 비용 × 검증 가능성으로 라우팅

14 · SOL’S FIRST JOB

Sol’s first job is not coding.
It is shaping the work.

01CLUSTER
Find independent work
→
02RISK
Attach failure cost
→
03MODEL
Choose execution tier
→
04EVIDENCE
Define proof

Route by failure cost × verifiability—not difficulty alone

15 · ROUTING MATRIX

모델 선택은
두 축이면 충분합니다.

실패 비용 ↑
자동 검증 가능성 →
SOL

모호하고 되돌리기 어려움

TERRA

중간 위험 · 구현 중심

LUNA

반복 많고 검증 쉬움

TERRA → SOL

실패 후 승격

15 · ROUTING MATRIX

Two axes are enough
for model routing.

FAILURE COST ↑
AUTOMATIC VERIFIABILITY →
SOL

Ambiguous + hard to reverse

TERRA

Medium risk · implementation

LUNA

High repetition · easy proof

TERRA → SOL

Escalate after failure

16 · PATTERN 1 · ONE SESSION

한 세션 안에서는
Sol만 Goal을 닫습니다.

SOLSHAPE · DELEGATE · INTEGRATE
TERRACOHERENT BUILD · ANALYSIS · VERIFY
LUNAHIGH-VOLUME · EASY-TO-CHECK

bounded task → artifact + verification + blocker · child마다 micro-Goal을 만들지 않습니다.

16 · PATTERN 1 · ONE SESSION

Inside one session,
only Sol closes the Goal.

SOLSHAPE · DELEGATE · INTEGRATE
TERRACOHERENT BUILD · ANALYSIS · VERIFY
LUNAHIGH-VOLUME · EASY-TO-CHECK

Bounded task → artifact + verification + blocker · no micro-Goal per child.

17 · GOALHOUSE

새 Goal은 빈 계획이 아니라
과거 episode에서 시작합니다.

BEFORE GOAL
  sync → recall nearest 3 episodes
  define gates + verifier
  choose PLANE 1 | PLANE 2
DURING
  append checkpoint + evidence locator
BEFORE COMPLETE
  검사맡기기 패킷
SOL ALONE CLOSES THE GOAL

한 가지 Taste · 원요청 → 변경점 → gate 증거 → 빠진 것

17 · GOALHOUSE

A new Goal starts from
a past episode—not a blank plan.

BEFORE GOAL
  sync → recall nearest 3 episodes
  define gates + verifier
  choose PLANE 1 | PLANE 2
DURING
  append checkpoint + evidence locator
BEFORE COMPLETE
  INSPECTION PACKET
SOL ALONE CLOSES THE GOAL

ONE TASTE · request → changes → gate evidence → gaps

18 · LIVE ARCHIFY · GOALHOUSE MAP

두 실행면을 직접 탐색하세요

드래그해서 이동 · 스크롤로 확대 · 노드를 눌러 관계 확인

INTERACTIVE HTML · ARCHIFY 2.14.0
전체 지도로 열기 ↗
18 · LIVE ARCHIFY · GOALHOUSE MAP

Explore the two execution planes

Drag to pan · scroll to zoom · select a node to inspect relationships

INTERACTIVE HTML · ARCHIFY 2.14.0
OPEN FULL MAP ↗
19 · PATTERN 2 · HOME / COMMAND CENTER

Command Center는 일하지 않고
맞는 세션을 오래 살립니다.

HOME · ROUTER ONLY
CLASSIFYDOMAIN × PRIVACY
RESOLVELIVE TASK STATE
HANDOFFLOCATOR + DONE_WHEN
WAITCOMPACT RECEIPT

실사용 · RALPHTHON DEPLOY HOME · 34 DAYS · 139 USER TURNS · 0 NATIVE GOAL · 혼합된 route는 격리

19 · PATTERN 2 · HOME / COMMAND CENTER

Command Center does not do the work.
It keeps the right task alive.

HOME · ROUTER ONLY
CLASSIFYDOMAIN × PRIVACY
RESOLVELIVE TASK STATE
HANDOFFLOCATOR + DONE_WHEN
WAITCOMPACT RECEIPT

IN USE · RALPHTHON DEPLOY HOME · 34 DAYS · 139 USER TURNS · 0 NATIVE GOAL · quarantine mixed routes

20 · SESSION PROTOCOL

세션 사이에는 대화가 아니라
다섯 필드의 계약을 전달합니다.

01OBJECTIVE02INPUTS03OUTPUT04PROOF05BLOCKER
OUTPUT: rendered HTML deck
PROOF: 34 slides × KO/EN, zero overflow
BLOCKER: missing publishable event photo
20 · SESSION PROTOCOL

Pass a five-field contract
between sessions—not conversation.

01OBJECTIVE02INPUTS03OUTPUT04PROOF05BLOCKER
OUTPUT: rendered HTML deck
PROOF: 34 slides × KO/EN, zero overflow
BLOCKER: missing publishable event photo
21 · PERFORMANCE PER VERIFIED ITERATION

가장 비싼 모델이 아니라
가장 싼 검증 가능한 루프

SOL은 선택합니다. TERRA는 완성합니다. LUNA는 반복량을 만듭니다.

ROUTE BY RISKVERIFY EVERYTHINGESCALATE ON FAILURESYNTHESIZE WITH SOL
21 · PERFORMANCE PER VERIFIED ITERATION

Not the strongest model—
the cheapest verifiable loop

SOL chooses. TERRA completes. LUNA creates iteration volume.

ROUTE BY RISKVERIFY EVERYTHINGESCALATE ON FAILURESYNTHESIZE WITH SOL
SESSION 03 · AFTER EXECUTION

Goal은 누구나 돌릴 수 있습니다.
이제 병목은 이해입니다.

LISTEN · ARGUE · SIMULATE · DECIDE

RESEARCHPODCASTCOMPANIONSIMULATION
SESSION 03 · AFTER EXECUTION

Anyone can run a Goal.
Understanding is the new bottleneck.

LISTEN · ARGUE · SIMULATE · DECIDE

RESEARCHPODCASTCOMPANIONSIMULATION
SESSION 03 · 25-MINUTE MAP

실행 이후의 병목,
이해를 설계하기

25분 뒤, 배우고 싶은 주제 하나를 듣고·논쟁하고·시뮬레이션하는 개인 학습 루프로 바꿀 수 있습니다.

25MINUTES
ONE OUTCOME
01 · MOVE THE BOTTLENECK실행에서 이해로

에이전트가 빨라질수록 무엇을 할지 판단하는 힘이 중요해집니다.

02 · BUILD INTERFACES다르게 부딪히기

논쟁형 Podcast · Founder Companion · Simulation

03 · DELEGATE RESEARCH연구 루프 살리기

Question → Hypothesis → Experiment → Evidence

TAKEAWAY · 한 주제를 Speak → Listen → Argue → Simulate로 설계

SESSION 03 · 25-MINUTE MAP

Design understanding—
the bottleneck after execution

After 25 minutes, you can turn one topic into a personal loop for listening, arguing, and simulating.

25MINUTES
ONE OUTCOME
01 · MOVE THE BOTTLENECKFrom execution to understanding

As agents get faster, deciding what deserves doing matters more.

02 · BUILD INTERFACESEncounter it differently

Adversarial podcast · Founder Companion · simulation

03 · DELEGATE RESEARCHKeep research alive

Question → Hypothesis → Experiment → Evidence

TAKEAWAY · Design one topic as Speak → Listen → Argue → Simulate

22 · MY PERSONAL HARNESS

제가 요즘 토큰을 가장 많이 쓰는 곳은
빌드가 아닙니다.

이해에 씁니다.

  1. 01나만의 논쟁형 팟캐스트를 만들어 듣기
  2. 02Codex + Realtime API로 에이전트와 회의하기
  3. 03Codex로 인터랙티브 시뮬레이션 만들어 틀려보기
PRIVATE PODCAST · RESEARCH TO LISTENING
PRIVATE PODCAST · RESEARCH TO LISTENING
CODEX + REALTIME · FOUNDER COMPANION
CODEX + REALTIME · FOUNDER COMPANION
CODEX · INTERACTIVE SIMULATION
CODEX · INTERACTIVE SIMULATION
22 · MY PERSONAL HARNESS

I spend most of my tokens
somewhere other than building.

I spend them on understanding.

  1. 01Generate an adversarial private podcast
  2. 02Meet an agent through Codex + Realtime API
  3. 03Build an interactive simulation and fail inside it
PRIVATE PODCAST · RESEARCH TO LISTENING
PRIVATE PODCAST · RESEARCH TO LISTENING
CODEX + REALTIME · FOUNDER COMPANION
CODEX + REALTIME · FOUNDER COMPANION
CODEX · INTERACTIVE SIMULATION
CODEX · INTERACTIVE SIMULATION
23 · RESEARCH → ARGUMENT

요약하지 말고,
반대자가 공격하게 합니다.

RESEARCH→SKEPTICAL HOST→SCRIPT→TTS→PODCAST
  1. 01근거가 충분한가?
  2. 02인과인가 상관인가?
  3. 03경제성이 있는가?
  4. 04어디서 깨지는가?
VERIFIED · 38:59 PRIVATE PODCAST
VERIFIED · 38:59 PRIVATE PODCAST
23 · RESEARCH → ARGUMENT

Do not summarize it.
Make a skeptic attack it.

RESEARCH→SKEPTICAL HOST→SCRIPT→TTS→PODCAST
  1. 01Is the evidence sufficient?
  2. 02Causality or correlation?
  3. 03Does the economics work?
  4. 04Where does it break?
VERIFIED · 38:59 PRIVATE PODCAST
VERIFIED · 38:59 PRIVATE PODCAST
24 · MEETING-SCOPED INTELLIGENCE

더 많은 메모리가 아니라,
이 회의에 필요한 정보만

Founder Companion은 판단을 대신하는 제품이 아니라 판단을 부딪히는 실험입니다.

EVIDENCEPRIOR DECISIONSHYPOTHESESCOUNTERARGUMENTSOPEN QUESTIONS
REALTIME FOUNDER COMPANION · IN USE
REALTIME FOUNDER COMPANION · IN USE
24 · MEETING-SCOPED INTELLIGENCE

Not more memory—
only what this meeting needs

Founder Companion is an experiment that challenges judgment—not a product that replaces it.

EVIDENCEPRIOR DECISIONSHYPOTHESESCOUNTERARGUMENTSOPEN QUESTIONS
REALTIME FOUNDER COMPANION · IN USE
REALTIME FOUNDER COMPANION · IN USE
25 · BUILD A WORLD

모르면 설명을 더 듣지 말고
작은 세계를 만듭니다.

STATE→CHOICE→CONSEQUENCE→REPLAY

보상을 바꿔보고 · reward hacking을 겪고 · 실패를 다시 플레이

INTERACTIVE LEARNING SOFTWARE
INTERACTIVE LEARNING SOFTWARE
25 · BUILD A WORLD

When explanation fails,
build a small world.

STATE→CHOICE→CONSEQUENCE→REPLAY

Change the reward · experience reward hacking · replay failure

INTERACTIVE LEARNING SOFTWARE
INTERACTIVE LEARNING SOFTWARE
26 · VALID FORMAT ≠ VALID JUDGMENT

JSON 정확도 100%
Decision 정확도 0%

학습 앱의 한 장면입니다. 벤치마크 수치가 아니라 차이를 몸으로 이해하는 장치입니다.

100 / 0
형식이 맞는 것과 판단이 맞는 것은 다릅니다.
PERSONAL LEARNING APP · NOT A BENCHMARK
26 · VALID FORMAT ≠ VALID JUDGMENT

100% JSON accuracy
0% decision accuracy

A scene from my learning app—not a benchmark, but a device for feeling the difference.

100 / 0
Valid format is not valid judgment.
PERSONAL LEARNING APP · NOT A BENCHMARK
27 · RALPHTHON @ ICML 2026

Task 위임 다음은
Innovation의 위임입니다.

하루 동안 AI가 질문·가설·실험·증거·수정을 어디까지 책임질 수 있는지 시험했습니다.

RALPHTHON @ ICML 2026
RALPHTHON @ ICML 2026
27 · RALPHTHON @ ICML 2026

After delegating tasks,
we delegate innovation.

For one day, we tested how much of question, hypothesis, experiment, evidence, and revision AI could own.

RALPHTHON @ ICML 2026
RALPHTHON @ ICML 2026
28 · A PROTOCOL THAT TRAVELS

Seoul → SF → Singapore →
Seoul again.

같은 이벤트를 반복한 것이 아니라, 같은 protocol을 다른 환경에 던졌습니다.

SEOUL
SEOUL
SAN FRANCISCO
SAN FRANCISCO
SINGAPORE
SINGAPORE
28 · A PROTOCOL THAT TRAVELS

Seoul → SF → Singapore →
Seoul again.

We did not repeat one event. We tested one protocol in different environments.

SEOUL
SEOUL
SAN FRANCISCO
SAN FRANCISCO
SINGAPORE
SINGAPORE
29 · THE LOBSTER RULE

Ralphthon이 시작되면
손을 떼세요.

농담 같은 규칙으로 구현을 다시 빼앗는 습관을 끊고, 사람은 목표·증거·안전 경계를 설계합니다.

노트북을 만지고 싶다면
lobster costume을 입으세요.
RALPHTHON @ ICML · ORIGINAL RULE
RALPHTHON @ ICML · ORIGINAL RULE
29 · THE LOBSTER RULE

When Ralphthon begins,
take your hands off.

A playful rule stops humans from taking implementation back. Humans design goals, evidence, and safety boundaries.

If you want to touch your laptop,
wear a lobster costume.
RALPHTHON @ ICML · ORIGINAL RULE
RALPHTHON @ ICML · ORIGINAL RULE
30 · WHY RESEARCH?

테스크 위임 다음은
혁신의 위임입니다.

TASK
정답이 있는 목표
완료 조건을 향한 loop

RESEARCH
정답이 없는 질문
가설·실험·증거가 다음 loop를 선택

TASK DELEGATION → INNOVATION DELEGATION
TASK DELEGATION → INNOVATION DELEGATION
30 · WHY RESEARCH?

After task delegation
comes innovation delegation.

TASK
A target with an answer
Loop toward completion

RESEARCH
A question without an answer
Hypothesis, experiment, evidence choose the next loop

TASK DELEGATION → INNOVATION DELEGATION
TASK DELEGATION → INNOVATION DELEGATION
31 · AUTO RESEARCH

Goal은 논문을 쓰게 하는 명령이 아니라
연구 루프를 살아 있게 하는 계약입니다.

실패도 다음 가설을 선택하게 만드는 정보라면 연구의 일부입니다.

QUESTION→HYPOTHESIS→EXPERIMENT→EVIDENCE→REVISION

실패도 다음 가설을 선택하게 만드는 정보라면 연구의 일부입니다.

ICML · GOAL IS A CONTRACT
ICML · GOAL IS A CONTRACT
31 · AUTO RESEARCH

Goal is not an order to write a paper.
It keeps the research loop alive.

Failure is part of research when it helps choose the next hypothesis.

QUESTION→HYPOTHESIS→EXPERIMENT→EVIDENCE→REVISION

Failure is part of research when it helps choose the next hypothesis.

ICML · GOAL IS A CONTRACT
ICML · GOAL IS A CONTRACT
32 · TOMORROW MORNING

한 주제를
네 개의 인터페이스로 공부하세요.

01말한다

정리되지 않은 생각을 꺼낸다

02듣는다

논쟁형 팟캐스트로 반복한다

03부딪힌다

Companion과 결정한다

04틀려본다

작은 세계에서 결과를 겪는다

32 · TOMORROW MORNING

Study one topic through
four interfaces.

01SPEAK

Externalize unstructured thought

02LISTEN

Repeat through adversarial audio

03ARGUE

Decide with a companion

04SIMULATE

Experience consequence in a small world

33 · THE HUMAN LOOP

에이전트가 실행을 가져갈수록,
인간은 더 깊이 이해해야 합니다.

GOAL
일을 끝냅니다.
UNDERSTANDING
어떤 일을 끝낼지 결정합니다.

Goal · Routing · Understanding은 하나의 운영 시스템입니다.

33 · THE HUMAN LOOP

As agents take execution,
humans must deepen understanding.

GOAL
Finishes the work.
UNDERSTANDING
Decides which work deserves finishing.

Goal · Routing · Understanding form one operating system.

01 / 42
← → · SPACE · K / E · F