Agents (LLM Agents)

Definition

LLM 을 핵심 controller 로 두고 multi-step planning + tool use + environment interaction 을 수행하는 시스템. 단일 응답을 넘어선 시퀀스적 작업 수행 능력.

대표 패턴:

• ReAct (Reason → Act → Observe 루프)
• Reflexion (자기 비판 + 재시도)
• Tool calling (function calling, MCP)
• Code generation as control flow

Why It Matters

• 본인 관심사 정중앙 (1.5x 가중치)
• 2026 AI 산업의 가장 두꺼운 흐름 — 거의 모든 frontier lab 이 agent 라인 보유
• LLM 의 "텍스트 생성기" → "범용 작업자" 전환점

State of the Art (2026-05)

산업 측

• Claude Opus 4.7 — Anthropic, agents/multi-step 강조
• GPT-Rosalind + Codex — OpenAI, OpenAI 의 software engineering agent
• Muse Spark (1.0 / 1.1) — Meta MSL, "personal" 강조
• Mistral Medium 3.5 + Vibe remote agents — 유럽 진영
• Google DeepMind AlphaEvolve — coding 에이전트

Agent Platforms & Infra (2026-06)

• Microsoft Build 2026 Agent Stack (2026-06-02 키노트 확정): 전체 에이전트 실행 스택 동시 출시.
  • Windows Agent Framework 1.0 — MIT 라이선스 오픈소스; Windows 11/365/Azure Arc 지원; human approval queue(권한 필요 작업 사람 승인 필수) 내장; 모델 무관(model-agnostic)
  • Azure Agent Mesh — on-prem Windows/Windows 365/Azure Arc edge에 걸친 federated 멀티에이전트 실행 플랫폼; 로컬과 동일 API 사용; 지연/GPU 가용성 기반 자동 라우팅; GA 목표 Q4 2026
  • GitHub Copilot App — 에이전트 네이티브 독립 데스크톱 앱 (macOS/Windows)
  • VS Code 멀티에이전트 GA — Build 당일
  • Copilot Workspace GA (Enterprise) — 버그 픽스/테스트 작성/PR 오픈 자율 에이전트
  • Azure AI Foundry — Claude, Mistral, Llama 4, DeepSeek 1차 지원 → 멀티모델 에이전트 오케스트레이션
  • Project Polaris — MoE 코딩 에이전트, GitHub Copilot 기본 모델 2026-08 GA
  • MAI-Code-1 / MAI-Code-1-Flash — 5B급, Copilot 전 티어 당일 GA
  • MAI-Thinking-1 — 35B active, 추론/오케스트레이션 특화 → Microsoft (source)

Non-Developer Agents: Codex for Every Role (2026-06-02)

OpenAI announced role-specific plugin expansion of Codex (source):

• 6 role-specific plugins with 62 popular apps + 110 skills (Analyst, Marketer, Operator, Designer, Researcher, Investor/Banker)
• Codex Sites (preview): creates interactive hosted web apps from text → dashboards, planners, review workspaces
• Non-developers: 20% of Codex users, growing 3× faster than developers
• Coming: Corporate Finance, PE, Marketing Strategy, Strategy Consulting, Legal plugins
• Significance: "AI agent" is crossing the developer ↔ non-developer boundary; software engineering agents generalizing to knowledge work. Most capable agent (Codex) now positioned as universal knowledge worker tool.

→ OpenAI

Multi-Agent Harness Design for Long-Running Tasks (Anthropic, April 2026 —

Anthropic Engineering Blog describes a three-agent architecture enabling multi-hour autonomous coding/frontend design sessions (source):

• Planner — expands a short product prompt into full spec
• Generator — builds the application
• Evaluator — uses Playwright MCP to test behavior against contracts (separating creation from critique)
• Context reset technique — completely clears context window at overflow, passes structured handoff (state + next steps) to fresh agent
• Results: solo run $9/20 min vs. full harness $200/6 hr → significantly more complete product
• Related: Anthropic 2026 Agentic Coding Trends Report (Jan 2026) — "delegation gap": developers delegate only 0-20% of tasks; harness design addresses quality problem of longer autonomous runs

Significance: Context reset + structured handoff is an emerging pattern for scaling agent runs beyond single context windows — enables production-quality outputs without manual re-intervention.

→ Anthropic

Computer Use (GUI Agents)

• Anthropic + Vercept (Feb 2026): Vercept 인수로 Claude computer use 강화. OSWorld <15% (2024-Q4) → 72.5% (2026-05). Cloud-hosted MacBook 원격 제어 아키텍처. → Anthropic (source)
• 브라우저·데스크톱 GUI 조작이 agents의 중요 실행 환경으로 부상 — 코드 실행 넘어 범용 컴퓨터 사용

연구 측

• Agentic Reinforcement Learning — RL 적 학습 방향
• Embodied Agents — 물리 세계 확장
• MCP (Model Context Protocol) — Anthropic 표준 → 광범위 채택 (Microsoft 365 Copilot 기본 통합 2026-01)

Google ADK 2.0 GA + Agents CLI (2026-06-30)

Google ADK 가 그래프 워크플로(fan-out/fan-in, 루프, 상태 관리, human-in-the-loop)와 에이전트 간 위임을 위한 협업 Task API 를 갖추고 GA 에 도달했다. 새로 나온 Agents CLI 는 도구 하나로 전 생애주기를 다룬다(스캐폴드 → 평가 → 배포 → 관측 → 퍼블리싱). Claude Code·Cursor·Gemini CLI 와 호환된다. 핵심 평가 공백을 겨냥한다 — 팀의 89% 가 관측성을 갖췄지만 평가 체계를 갖춘 곳은 52% 에 불과하다. → Google ADK (Agent Development Kit), Google DeepMind (source)

Sub-concepts

• Agentic Reinforcement Learning — RL 측면
• Embodied Agents — 물리 세계 측면
• tool-use, mcp, reflection 별도 페이지 TBD

Persistent Memory: OpenAI Dreaming V3 (2026-06-04)

OpenAI released Dreaming V3 — a background memory synthesis architecture that represents the most significant advance in agent/user memory since the original ChatGPT memory rollout (source):

• Mechanism: Background process continuously synthesizes salient facts, preferences, and temporal context from all conversations — no explicit user instruction required
• Temporal awareness: Auto-updates time-sensitive memories (e.g., past events are revised from future tense to past tense)
• Replaces: The explicit "saved memories" list as the standalone memory foundation
• Performance: Factual recall 41.5% (2024) → 82.8% (2026); preference/time-sensitive accuracy in low-70s
• Compute: 5× reduction in memory-related inference compute → enables Free-tier rollout
• Rollout: June 4, Plus/Pro US first; Free and global to follow

Disambiguation: OpenAI "Dreaming" (personalization memory for ChatGPT users) vs Anthropic "Dreaming" (agents reviewing past sessions to form procedural memory for self-improvement — see Claude Managed Agents). Same name, different mechanism.

Agent relevance: Persistent user memory is a foundational requirement for continuity in long-running agents. Dreaming V3 solves this at the product layer for consumer ChatGPT; the architectural pattern (background synthesis vs. explicit storage) is the key technique.

→ OpenAI

Open Problems

• Long-horizon stability — 50+ step 작업에서 drift / 실패 누적
• Verification — agent 가 자기 작업 결과 검증 가능?
• Memory — context 한계 vs 누적 학습 (Dreaming V3가 소비자 레이어 해법 제시)
• Cost — agent run 당 token 비용 vs gain
• Safety — autonomous action 의 부작용

Key Papers / Events

• Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent — 2026-06-30: 에이전틱 과제에서는 horizon 스케일링이 파라미터 스케일링을 앞선다 (45K 토큰 궤적으로 35B MoE 가 1T급 모델과 대등)
• Self-Distilled Agentic Reinforcement Learning — 2026-05 latest

Related Concepts

• Agentic Reinforcement Learning
• Embodied Agents
• Reasoning Models

Notable Statements

• Andrej Karpathy (2026): "80% manual + 20% agents 에서 80% agent coding + 20% edits 로 빠르게 전환"
• Jim Fan: "LLM acts as 'prefrontal cortex' that orchestrates lower-level control APIs"
• Andrej Karpathy: "creating software at the end of 2026 어떻게 보일지 상상하기 어렵다" (via Sam Altman echo)

Open Debates

• Agent = LLM + scaffolding vs Agent 는 별도 학습 필요 — 학계 일부 (agentic RL 진영) 는 후자 주장. 산업 일부는 전자로 충분 주장.
• No-gradient orchestration (Jim Fan 입장) vs end-to-end neural control (Sergey Levine 진영) — 학습 패러다임 갈림.