Skip to content

Latest commit

Β 

History

History
373 lines (250 loc) Β· 16.1 KB

File metadata and controls

373 lines (250 loc) Β· 16.1 KB

Workflow Reference β€” 6-Stage Pipeline

This document describes each stage of the /coord pipeline in detail, including inputs, processing logic, outputs, and the tools involved.

이 λ¬Έμ„œλŠ” /coord νŒŒμ΄ν”„λΌμΈμ˜ 각 단계λ₯Ό μž…λ ₯, 처리 둜직, 좜λ ₯, κ΄€λ ¨ 도ꡬλ₯Ό ν¬ν•¨ν•˜μ—¬ μƒμ„Ένžˆ μ„€λͺ…ν•©λ‹ˆλ‹€.


Pipeline Overview (νŒŒμ΄ν”„λΌμΈ κ°œμš”)

User message
     β”‚
     β–Ό
 Stage 0: BOOTSTRAP
     β”‚  Complexity score > threshold?
     β”‚  Yes ──────────────────────────►  Enter 6-stage pipeline
     β”‚  No  ──────────────────────────►  Direct native response
     β–Ό
 Stage 1: MEMORY PRE-CHECK  (parallel / 병렬)
     β”‚  Auto-memory  β€–  Serena  β€–  Graphiti
     β–Ό
 Stage 2: PLAN
     β”‚  Sequential-thinking MCP β†’ dependency map β†’ parallelization analysis
     β–Ό
 Stage 3: RESEARCH  (parallel where independent / 독립 μž‘μ—…μ€ 병렬)
     β”‚  Explore agents  β€–  deep-research-agent  β€–  context7  β€–  tavily
     β–Ό
 Stage 4: EXECUTE
     β”‚  Specialist agents β†’ intermediate lint/test β†’ TodoWrite tracking
     β–Ό
 Stage 5: REVIEW
     β”‚  critic agent (parallel with generator if Opus model)
     β”‚  Optional: premium model review via OpenRouter
     β–Ό
 Stage 6: MEMORY UPDATE
     β”‚  Route lessons β†’ Auto-memory > Graphiti > Serena
     β”‚  Update MEMORY.md index
     β–Ό
   Done

Stage 0 β€” Bootstrap (μ€€λΉ„ 단계)

Purpose: Measure task complexity and decide whether to engage the full pipeline or respond directly.

λͺ©μ : μž‘μ—… λ³΅μž‘λ„λ₯Ό μΈ‘μ •ν•˜κ³  전체 νŒŒμ΄ν”„λΌμΈμ„ μ‹€ν–‰ν• μ§€ 직접 응닡할지 κ²°μ •ν•©λ‹ˆλ‹€.

Input (μž…λ ₯)

  • Raw user message text (μ›μ‹œ μ‚¬μš©μž λ©”μ‹œμ§€ ν…μŠ€νŠΈ)
  • Active Claude Code session context (ν™œμ„± Claude Code μ„Έμ…˜ μ»¨ν…μŠ€νŠΈ)

Processing (처리)

The bootstrap stage scores the task across four dimensions:

Dimension (차원) Signal (μ‹ ν˜Έ) Weight (κ°€μ€‘μΉ˜)
Scope ambiguity (λ²”μœ„ λͺ¨ν˜Έμ„±) Vague verbs, qualifiers like "maybe", "explore" 0.25
Domain count (도메인 수) Mentions of frontend + backend, infra + code, etc. 0.30
Step count (단계 수) Estimated number of sequential operations 0.25
Risk level (μœ„ν—˜ μˆ˜μ€€) Production environment, destructive ops, PII 0.20

If the composite score exceeds 0.5, the full pipeline activates. Below that threshold, Claude responds without the pipeline overhead.

볡합 μ μˆ˜κ°€ 0.5λ₯Ό μ΄ˆκ³Όν•˜λ©΄ 전체 νŒŒμ΄ν”„λΌμΈμ΄ ν™œμ„±ν™”λ©λ‹ˆλ‹€. κ·Έ μ΄ν•˜μ—μ„œλŠ” νŒŒμ΄ν”„λΌμΈ μ˜€λ²„ν—€λ“œ 없이 Claudeκ°€ 직접 μ‘λ‹΅ν•©λ‹ˆλ‹€.

Output (좜λ ₯)

  • mode: pipeline or direct (νŒŒμ΄ν”„λΌμΈ λ˜λŠ” 직접)
  • complexity_score: float 0.0–1.0
  • task_summary: one-sentence restatement of the request (μš”μ²­μ˜ ν•œ λ¬Έμž₯ μž¬μ§„μˆ )

Stage 1 β€” Memory Pre-Check (λ©”λͺ¨λ¦¬ 사전 확인)

Purpose: Retrieve all relevant prior context before any planning occurs, so the plan does not repeat solved problems.

λͺ©μ : κ³„νšμ΄ λ°œμƒν•˜κΈ° 전에 λͺ¨λ“  κ΄€λ ¨ 이전 μ»¨ν…μŠ€νŠΈλ₯Ό κ²€μƒ‰ν•˜μ—¬ κ³„νšμ΄ 이미 ν•΄κ²°λœ 문제λ₯Ό λ°˜λ³΅ν•˜μ§€ μ•Šλ„λ‘ ν•©λ‹ˆλ‹€.

Input (μž…λ ₯)

  • task_summary from Stage 0 (Stage 0의 μž‘μ—… μš”μ•½)
  • Known project identifiers (μ•Œλ €μ§„ ν”„λ‘œμ νŠΈ μ‹λ³„μž)

Processing β€” Parallel scan (병렬 μŠ€μΊ”)

All three sources are queried simultaneously (μ„Έ μ†ŒμŠ€λ₯Ό λ™μ‹œμ— μ‘°νšŒν•©λ‹ˆλ‹€):

Source 1: Auto-memory (~/.claude/projects/…/memory/)

scan MEMORY.md index
└── read relevant topic files: feedback_*.md, *_project.md, lessons/

Source 2: Serena MCP

list_memories()          β†’ active workspace memory
read_memory("*_plan")   β†’ any existing implementation plans

Source 3: Graphiti (graphiti-memory)

search_nodes(query=task_summary)        β†’ related entities
search_memory_facts(query=task_summary) β†’ relationships and decisions

Note: Graphiti requires an OpenRouter embedding API key. If unavailable, this source is skipped gracefully and a warning is logged. (GraphitiλŠ” OpenRouter μž„λ² λ”© API ν‚€κ°€ ν•„μš”ν•©λ‹ˆλ‹€. μ‚¬μš©ν•  수 μ—†λŠ” 경우 κ²½κ³ λ₯Ό κΈ°λ‘ν•˜κ³  κ±΄λ„ˆλœλ‹ˆλ‹€.)

Output (좜λ ₯)

  • prior_context: merged dict of all retrieved facts (κ²€μƒ‰λœ λͺ¨λ“  μ‚¬μ‹€μ˜ 병합 λ”•μ…”λ„ˆλ¦¬)
  • lessons_loaded: list of lesson file names that were read (읽은 κ΅ν›ˆ 파일 이름 λͺ©λ‘)
  • existing_plan: optional, if a plan was found in memory (선택적, λ©”λͺ¨λ¦¬μ—μ„œ κ³„νšμ΄ 발견된 경우)

Stage 2 β€” Plan (κ³„νš 단계)

Purpose: Produce a structured, dependency-aware execution plan that explicitly identifies which steps can run in parallel.

λͺ©μ : λ³‘λ ¬λ‘œ μ‹€ν–‰ν•  수 μžˆλŠ” 단계λ₯Ό λͺ…μ‹œμ μœΌλ‘œ μ‹λ³„ν•˜λŠ” κ΅¬μ‘°ν™”λœ μ˜μ‘΄μ„± 인식 μ‹€ν–‰ κ³„νšμ„ μƒμ„±ν•©λ‹ˆλ‹€.

Input (μž…λ ₯)

  • task_summary (μž‘μ—… μš”μ•½)
  • prior_context from Stage 1 (Stage 1의 이전 μ»¨ν…μŠ€νŠΈ)

Processing (처리)

The sequential-thinking MCP is the primary tool for this stage. It performs structured multi-step reasoning with hypothesis testing.

sequential-thinking MCPκ°€ 이 λ‹¨κ³„μ˜ κΈ°λ³Έ λ„κ΅¬μž…λ‹ˆλ‹€. κ°€μ„€ ν…ŒμŠ€νŠΈμ™€ ν•¨κ»˜ κ΅¬μ‘°ν™”λœ 닀단계 좔둠을 μˆ˜ν–‰ν•©λ‹ˆλ‹€.

sequential-thinking:
  step 1: Understand the task boundaries (μž‘μ—… 경계 이해)
  step 2: Identify all required subtasks (ν•„μš”ν•œ λͺ¨λ“  ν•˜μœ„ μž‘μ—… 식별)
  step 3: Map dependencies between subtasks (ν•˜μœ„ μž‘μ—… κ°„ μ˜μ‘΄μ„± λ§€ν•‘)
  step 4: Separate sequential chains from parallel groups (순차 체인과 병렬 κ·Έλ£Ή 뢄리)
  step 5: Estimate resource usage and agent types needed (λ¦¬μ†ŒμŠ€ μ‚¬μš©λŸ‰ 및 ν•„μš”ν•œ μ—μ΄μ „νŠΈ μœ ν˜• μΆ”μ •)
  step 6: Output structured plan with parallelization annotations (병렬화 주석과 ν•¨κ»˜ κ΅¬μ‘°ν™”λœ κ³„νš 좜λ ₯)

If the task is ambiguous, a requirements-analyst subagent clarifies the request before planning continues.

μž‘μ—…μ΄ λͺ¨ν˜Έν•œ 경우, requirements-analyst μ„œλΈŒμ—μ΄μ „νŠΈκ°€ κ³„νšμ„ κ³„μ†ν•˜κΈ° 전에 μš”μ²­μ„ λͺ…ν™•νžˆ ν•©λ‹ˆλ‹€.

Parallelization principle (병렬화 원칙)

The plan explicitly marks every task as one of:

κ³„νšμ€ λͺ¨λ“  μž‘μ—…μ„ λ‹€μŒ 쀑 ν•˜λ‚˜λ‘œ λͺ…μ‹œμ μœΌλ‘œ ν‘œμ‹œν•©λ‹ˆλ‹€:

  • PARALLEL β€” can start at the same time as other PARALLEL tasks (λ‹€λ₯Έ PARALLEL μž‘μ—…κ³Ό λ™μ‹œμ— μ‹œμž‘ κ°€λŠ₯)
  • SEQUENTIAL(after=X) β€” must wait for task X to complete (μž‘μ—… Xκ°€ μ™„λ£Œλ  λ•ŒκΉŒμ§€ κΈ°λ‹€λ €μ•Ό 함)
  • BLOCKING β€” must complete before any next stage begins (λ‹€μŒ 단계가 μ‹œμž‘λ˜κΈ° 전에 μ™„λ£Œλ˜μ–΄μ•Ό 함)

Why parallelization matters: Running 3 independent file reads in parallel saves 60–70% of wall-clock time compared to sequential execution. The plan stage forces this analysis upfront so execution does not default to unnecessary serialization.

병렬화가 μ€‘μš”ν•œ 이유: 3개의 독립적인 파일 읽기λ₯Ό λ³‘λ ¬λ‘œ μ‹€ν–‰ν•˜λ©΄ 순차 싀행에 λΉ„ν•΄ μ‹€μ œ μ‹œκ°„μ˜ 60-70%λ₯Ό μ ˆμ•½ν•  수 μžˆμŠ΅λ‹ˆλ‹€.

Output (좜λ ₯)

  • Structured plan document written to write_memory("current_plan", …) (κ΅¬μ‘°ν™”λœ κ³„νš λ¬Έμ„œ)
  • TodoWrite tasks created for all subtasks (λͺ¨λ“  ν•˜μœ„ μž‘μ—…μ— λŒ€ν•œ TodoWrite μž‘μ—… 생성)

Stage 3 β€” Research (λ¦¬μ„œμΉ˜ 단계)

Purpose: Gather all information required for execution before any code is written or files are modified.

λͺ©μ : μ½”λ“œκ°€ μž‘μ„±λ˜κ±°λ‚˜ 파일이 μˆ˜μ •λ˜κΈ° 전에 싀행에 ν•„μš”ν•œ λͺ¨λ“  정보λ₯Ό μˆ˜μ§‘ν•©λ‹ˆλ‹€.

Input (μž…λ ₯)

  • Plan from Stage 2 (Stage 2의 κ³„νš)
  • Research tasks identified in the plan (κ³„νšμ—μ„œ μ‹λ³„λœ λ¦¬μ„œμΉ˜ μž‘μ—…)

Processing β€” Parallel agent dispatch (병렬 μ—μ΄μ „νŠΈ λ°œμ†‘)

Multiple research subagents are dispatched simultaneously. Each covers a different domain.

μ—¬λŸ¬ λ¦¬μ„œμΉ˜ μ„œλΈŒμ—μ΄μ „νŠΈκ°€ λ™μ‹œμ— λ°œμ†‘λ©λ‹ˆλ‹€. 각각 λ‹€λ₯Έ 도메인을 λ‹΄λ‹Ήν•©λ‹ˆλ‹€.

Parallel dispatch (병렬 λ°œμ†‘):
β”œβ”€β”€ Explore agent(s)         β†’ codebase structure, existing patterns
β”œβ”€β”€ deep-research-agent      β†’ web sources, academic papers, blog posts
└── context7 MCP             β†’ official library documentation

Additional routing:
β”œβ”€β”€ tavily                   β†’ current events, recent releases
└── notebooklm               β†’ long-form document analysis

Research routing rules (λ¦¬μ„œμΉ˜ λΌμš°νŒ… κ·œμΉ™):

Query type (쿼리 μœ ν˜•) Route to (λΌμš°νŒ… λŒ€μƒ)
"How does X work in the codebase?" Explore agent
"What is the latest version of X?" Tavily
"How does library X handle Y?" Context7 MCP
"Summarize this 50-page PDF" NotebookLM
"Compare three approaches to Z" deep-research-agent

Output (좜λ ₯)

  • research_findings: merged summary from all agents (λͺ¨λ“  μ—μ΄μ „νŠΈμ˜ 병합 μš”μ•½)
  • Source citations with credibility scores (신뒰도 μ μˆ˜κ°€ μžˆλŠ” μ†ŒμŠ€ 인용)

Stage 4 β€” Execute (μ‹€ν–‰ 단계)

Purpose: Implement the plan using specialist agents, with intermediate validation gates.

λͺ©μ : 쀑간 검증 κ²Œμ΄νŠΈμ™€ ν•¨κ»˜ μ „λ¬Έκ°€ μ—μ΄μ „νŠΈλ₯Ό μ‚¬μš©ν•˜μ—¬ κ³„νšμ„ κ΅¬ν˜„ν•©λ‹ˆλ‹€.

Input (μž…λ ₯)

  • Approved plan (승인된 κ³„νš)
  • Research findings (λ¦¬μ„œμΉ˜ κ²°κ³Ό)
  • TodoWrite task list (μž‘μ—… λͺ©λ‘)

Processing (처리)

Tasks are dispatched to specialist agents matching the work type. Agents update TodoWrite status as they complete each item.

μž‘μ—…μ€ μž‘μ—… μœ ν˜•μ— λ§žλŠ” μ „λ¬Έκ°€ μ—μ΄μ „νŠΈμ—κ²Œ λ°œμ†‘λ©λ‹ˆλ‹€. μ—μ΄μ „νŠΈλŠ” 각 ν•­λͺ©μ„ μ™„λ£Œν•  λ•Œ TodoWrite μƒνƒœλ₯Ό μ—…λ°μ΄νŠΈν•©λ‹ˆλ‹€.

Agent selection by task type (μž‘μ—… μœ ν˜•λ³„ μ—μ΄μ „νŠΈ 선택):

Task (μž‘μ—…) Agent (μ—μ΄μ „νŠΈ)
API design, server logic backend-architect
React components, CSS frontend-architect
Data pipelines, scripts python-expert
Full system cross-domain system-architect
Security audit security-engineer
Performance bottleneck performance-engineer
Test suite creation quality-engineer
Docs and guides technical-writer
Code smells, refactor refactoring-expert

Intermediate validation (쀑간 검증):

After each logical group of changes, the pipeline runs:

각 논리적 λ³€κ²½ κ·Έλ£Ή ν›„ νŒŒμ΄ν”„λΌμΈμ΄ μ‹€ν–‰ν•©λ‹ˆλ‹€:

# Language-appropriate checks (언어에 λ”°λ₯Έ 검사)
python -m pytest                   # Python projects
npm run lint && npm run typecheck  # JavaScript/TypeScript
go test ./...                      # Go projects
cargo test                         # Rust projects

Failures block further execution and trigger a root-cause-analyst agent.

μ‹€νŒ¨λŠ” μΆ”κ°€ 싀행을 μ°¨λ‹¨ν•˜κ³  root-cause-analyst μ—μ΄μ „νŠΈλ₯Ό νŠΈλ¦¬κ±°ν•©λ‹ˆλ‹€.

Output (좜λ ₯)

  • All planned code changes applied and validated (κ³„νšλœ λͺ¨λ“  μ½”λ“œ λ³€κ²½ 적용 및 검증)
  • Updated TodoWrite with completed statuses (μ™„λ£Œ μƒνƒœκ°€ μ—…λ°μ΄νŠΈλœ TodoWrite)
  • Memory checkpoint: write_memory("checkpoint_…", current_state) (λ©”λͺ¨λ¦¬ 체크포인트)

Stage 5 β€” Review (리뷰 단계)

Purpose: Independent quality check to catch errors, compliance issues, and sycophantic reasoning in the output.

λͺ©μ : 좜λ ₯μ—μ„œ 였λ₯˜, κ·œμ • μ€€μˆ˜ 문제, 아첨적 좔둠을 λ°œκ²¬ν•˜κΈ° μœ„ν•œ 독립적인 ν’ˆμ§ˆ κ²€μ‚¬μž…λ‹ˆλ‹€.

Input (μž…λ ₯)

  • All artifacts produced in Stage 4 (Stage 4μ—μ„œ μƒμ„±λœ λͺ¨λ“  κ²°κ³Όλ¬Ό)
  • Original plan and requirements (μ›λž˜ κ³„νš 및 μš”κ΅¬μ‚¬ν•­)

Processing (처리)

Standard review (ν‘œμ€€ 리뷰):

The critic agent runs in parallel with the primary generator when an Opus model is used. It applies Popperian falsifiability checks: every claim must include a "wrong if X" condition.

critic μ—μ΄μ „νŠΈλŠ” Opus λͺ¨λΈ μ‚¬μš© μ‹œ κΈ°λ³Έ 생성기와 λ³‘λ ¬λ‘œ μ‹€ν–‰λ©λ‹ˆλ‹€. Popperian 반증 κ°€λŠ₯μ„± 검사λ₯Ό μ μš©ν•©λ‹ˆλ‹€: λͺ¨λ“  μ£Όμž₯μ—λŠ” "X인 경우 ν‹€λ¦Ό" 쑰건이 ν¬ν•¨λ˜μ–΄μ•Ό ν•©λ‹ˆλ‹€.

Premium review (프리미엄 리뷰, optional):

For high-stakes work (proposals, papers, production deployments), a second independent review uses a premium model via OpenRouter:

κ³ μœ„ν—˜ μž‘μ—…(μ œμ•ˆμ„œ, λ…Όλ¬Έ, ν”„λ‘œλ•μ…˜ 배포)의 경우, 두 번째 독립적인 κ²€ν† κ°€ OpenRouterλ₯Ό ν†΅ν•œ 프리미엄 λͺ¨λΈμ„ μ‚¬μš©ν•©λ‹ˆλ‹€:

OpenRouter β†’ Gemini 2.5 Pro  (κ²€ν†  1)
OpenRouter β†’ GPT-4.1 o3     (κ²€ν†  2)

Output (좜λ ₯)

  • Review report with pass/fail per criterion (기쀀별 톡과/μ‹€νŒ¨κ°€ μžˆλŠ” κ²€ν†  λ³΄κ³ μ„œ)
  • Falsifiability score: fraction of claims with wrong-if-X conditions (반증 κ°€λŠ₯μ„± 점수)
  • Approval to proceed, or revision requests (μ§„ν–‰ 승인 λ˜λŠ” μˆ˜μ • μš”μ²­)

Stage 6 β€” Memory Update (λ©”λͺ¨λ¦¬ μ—…λ°μ΄νŠΈ 단계)

Purpose: Preserve lessons, decisions, and new knowledge for future sessions.

λͺ©μ : 미래 μ„Έμ…˜μ„ μœ„ν•΄ κ΅ν›ˆ, κ²°μ •, μƒˆλ‘œμš΄ 지식을 λ³΄μ‘΄ν•©λ‹ˆλ‹€.

Input (μž…λ ₯)

  • Session outcomes (μ„Έμ…˜ κ²°κ³Ό)
  • Review findings (κ²€ν†  κ²°κ³Ό)
  • Any errors or surprises encountered (λ°œμƒν•œ 였λ₯˜ λ˜λŠ” μ˜ˆμƒμΉ˜ λͺ»ν•œ 상황)

Memory routing rules (λ©”λͺ¨λ¦¬ λΌμš°νŒ… κ·œμΉ™)

Content type (λ‚΄μš© μœ ν˜•) Route to (λΌμš°νŒ… λŒ€μƒ) File pattern (파일 νŒ¨ν„΄)
Code locations, symbols (μ½”λ“œ μœ„μΉ˜, 심볼) Serena MCP workspace memory
User preferences, lessons (μ‚¬μš©μž μ„ ν˜Έ, κ΅ν›ˆ) Auto-memory feedback_*.md
Project status (ν”„λ‘œμ νŠΈ μƒνƒœ) Auto-memory *_project.md
Domain facts, decisions (도메인 사싀, κ²°μ •) Graphiti (if running) knowledge graph node
Session summary (μ„Έμ…˜ μš”μ•½) Auto-memory *_handoff.md

Priority order (μš°μ„ μˆœμœ„ μˆœμ„œ): Auto-memory > Graphiti > Serena

Processing (처리)

1. write_memory("session_summary", outcomes)         β†’ Serena
2. Append lesson to feedback_*.md                    β†’ Auto-memory
3. Update MEMORY.md index (one line per entry)       β†’ Auto-memory
4. graphiti: add_memory(episode) for domain facts    β†’ Graphiti (if available)
5. delete_memory("checkpoint_*") for temp items      β†’ Serena

Output (좜λ ₯)

  • Updated persistent memory across all three stores (μ„Έ μ €μž₯μ†Œ λͺ¨λ‘μ˜ μ—…λ°μ΄νŠΈλœ 영ꡬ λ©”λͺ¨λ¦¬)
  • MEMORY.md index reflects new entries (MEMORY.md μΈλ±μŠ€κ°€ μƒˆ ν•­λͺ©μ„ 반영)

Anti-Patterns (ν”Όν•΄μ•Ό ν•  νŒ¨ν„΄)

The following patterns violate the pipeline's design principles and will produce worse outcomes than simply responding directly.

λ‹€μŒ νŒ¨ν„΄μ€ νŒŒμ΄ν”„λΌμΈμ˜ 섀계 원칙을 μœ„λ°˜ν•˜λ©° 직접 μ‘λ‹΅ν•˜λŠ” 것보닀 더 λ‚˜μœ κ²°κ³Όλ₯Ό λ‚³μŠ΅λ‹ˆλ‹€.

Anti-pattern (ν”Όν•΄μ•Ό ν•  νŒ¨ν„΄) Why it is harmful (ν•΄λ‘œμš΄ 이유)
Skipping memory check (λ©”λͺ¨λ¦¬ 확인 κ±΄λ„ˆλ›°κΈ°) Repeats solved problems from previous sessions (이전 μ„Έμ…˜μ—μ„œ ν•΄κ²°λœ 문제 반볡)
Jump-to-implementation (κ³„νš 없이 μ¦‰μ‹œ κ΅¬ν˜„) Produces code that misses edge cases and requirements (μ—£μ§€ μΌ€μ΄μŠ€ 및 μš”κ΅¬μ‚¬ν•­μ„ λ†“μΉœ μ½”λ“œ 생성)
Sequential where parallel is possible (병렬 κ°€λŠ₯ μ‹œ 순차 μ‹€ν–‰) Wastes 60–70% of available execution time (μ‚¬μš© κ°€λŠ₯ν•œ μ‹€ν–‰ μ‹œκ°„μ˜ 60-70% λ‚­λΉ„)
Using /coord for simple tasks (λ‹¨μˆœ μž‘μ—…μ— /coord μ‚¬μš©) Pipeline overhead exceeds task time (νŒŒμ΄ν”„λΌμΈ μ˜€λ²„ν—€λ“œκ°€ μž‘μ—… μ‹œκ°„ 초과)
Skipping critic review (critic κ²€ν†  κ±΄λ„ˆλ›°κΈ°) Sycophantic reasoning goes undetected (아첨적 좔둠이 κ°μ§€λ˜μ§€ μ•ŠμŒ)
Omitting memory update (λ©”λͺ¨λ¦¬ μ—…λ°μ΄νŠΈ μƒλž΅) Lessons are lost and re-learned unnecessarily (κ΅ν›ˆμ΄ 사라지고 λΆˆν•„μš”ν•˜κ²Œ λ‹€μ‹œ ν•™μŠ΅)
3+ sequential searches in one context Delegate to parallel subagents instead (병렬 μ„œλΈŒμ—μ΄μ „νŠΈμ— μœ„μž„)