티스토리 뷰

Engineering

Prime Agent 와 Hermes Agent 비교

이도운 2026. 8. 26. 22:28
Hermes Agent는 Skill을 스스로 만들고 개선하는 Agent Framework로, OpenClaw의 대체로 사용자층을 많이 확보해오고 있습니다. Coding harness로도 많이 사용되고 있는데요, 얼마 전 조금 다른 방식으로 스스로 개선한다는 Prime Agent가 소개되어 서로 비교하는 문서를 만들어 봤습니다. 바쁘시면 15번부터 보세요.

1. 개요

Prime Agent는 Prime Intellect가 2026년 8월 5일 공개한 오픈소스 Coding/Research Agent이다. 기존 Tool Calling 중심 Agent Harness와 달리 Recursive Language Model(RLM)Continual Harness를 주요 구조로 사용한다. 2026년 8월 24일에는 관련 기술 보고서도 공개되었다.

Hermes Agent는 Nous Research가 개발하는 범용 Autonomous Agent로, Persistent Memory와 Skill을 중심으로 한 Closed Learning Loop를 제공한다. 대화 및 작업 경험을 Memory와 Skill로 축적하고 이후 Session에서 재사용하는 것을 주요 특징으로 한다.

두 Agent 모두 "Self-Improving Agent"를 표방하지만 자기개선 대상으로 보는 범위와 실행 구조에는 차이가 있다.


2. Prime Agent

2.1 주요 목적

Prime Agent는 다음 작업을 주요 대상으로 한다.

  • Coding
  • Research
  • Long-running Task
  • Autonomous Evaluation
  • Multi-Agent 작업
  • 반복적인 Agent Workflow

일반적인 질의응답형 Agent보다는 장시간 지속되는 복합 작업을 전제로 설계되어 있다.


2.2 핵심 구조

Prime Agent는 다음 두 구조를 중심으로 한다.

Prime Agent
├── Recursive Language Model
│   ├── Persistent IPython
│   ├── Programmatic Tool Call
│   ├── Context Processing
│   └── Recursive Subagent
│
└── Continual Harness
    ├── Prompt
    ├── Memory
    ├── Skill
    └── Subagent Specification

Recursive Language Model

RLM은 LLM이 Tool을 개별 Function Calling Schema 형태로 직접 호출하기보다 Persistent Python 환경을 통해 프로그램 방식으로 작업을 구성하도록 하는 구조이다.

Prime Agent에서 모델에 기본적으로 제공되는 Tool은 Persistent IPython 환경이며 다음 작업들이 Python 코드 형태로 수행된다.

  • 파일 처리
  • Shell 실행
  • Tool 호출
  • Context 처리
  • Subagent 실행
  • 반복 및 조건 처리
  • 결과 집계

일반적인 Tool Calling 구조는 다음과 같다.

LLM
 ├── Search Tool
 ├── File Tool
 ├── Shell Tool
 ├── MCP Tool
 └── Delegate Tool

Prime Agent에서는 다음과 같은 구조에 가깝다.

LLM
 │
 └── Persistent IPython
       ├── Python
       ├── File
       ├── Shell
       ├── MCP
       ├── Skill
       └── rlm()
             └── Subagent

Tool 호출뿐 아니라 Tool 간 조합과 제어 흐름도 LLM이 생성한 프로그램을 통해 처리할 수 있다.


2.3 Recursive Subagent

Prime Agent에서는 rlm(...)을 이용하여 Child Agent를 생성할 수 있다.

Parent Agent
   │
   ├── rlm() → Agent A
   ├── rlm() → Agent B
   └── rlm() → Agent C

Subagent는 별도의 Prime Agent Instance로 실행되며 병렬 작업 및 Background 작업에 사용할 수 있다.

또한 실행 중인 Agent 간 직접 메시지 교환을 지원한다.

Agent A ←→ Agent B
   ↑          ↑
   └─ Parent ─┘

모든 Agent 간 통신을 사용자 또는 중앙 Supervisor가 중계해야 하는 구조는 아니다.


3. Continual Harness

3.1 개념

Continual Harness는 Agent가 작업 과정에서 획득한 정보를 현재 Conversation Context 외부의 지속 가능한 상태로 보관하는 구조이다.

주요 관리 대상은 다음과 같다.

 

대상 내용
Prompt Base System Prompt 이외의 보조 지침
Memory 사실, 결정, 실패, 선호, 결과
Skill 재사용 가능한 작업 방법
Subagent Spec 재사용 가능한 Subagent 정의

따라서 Prime Agent의 지속 상태는 일반적인 Memory보다 범위가 넓다.

Persistent Agent State
    │
    ├── Prompt
    ├── Memory
    ├── Skill
    └── Agent Configuration

3.2 Refinement

Prime Agent는 /refine을 통해 현재 작업 Trajectory를 검토하고 Continual Harness를 수정할 수 있다.

Execution Trajectory
        │
        ▼
     /refine
        │
        ▼
Evidence-based Update
        │
        ├── Prompt
        ├── Memory
        ├── Skill Description
        └── Subagent Spec

Refinement에는 다음 제한이 있다.

  • Base System Prompt는 변경하지 않음
  • Supplemental Harness State만 변경
  • 변경 이력을 기록
  • Snapshot을 이용한 Rollback 지원
  • 기본적으로 Session-local 범위에서 동작

주의할 점은 /refine이 자동으로 모든 Executable Skill을 생성하는 기능과 동일하지 않다는 것이다.

Prime Agent의 Executable Skill은 Import 가능한 Python Package 형태이며, 별도의 Skill Creator를 통해 생성할 수 있다. 공식 문서에서도 Continual Harness Refinement가 새로운 Executable Skill의 Packaging 및 Review를 대체하지 않는다고 명시하고 있다.


4. Long-running Agent 기능

Prime Agent는 장시간 실행을 위해 다음 기능을 제공한다.

  • Persistent Goal
  • Background Daemon
  • Session Detach / Reattach
  • Heartbeat
  • Schedule
  • Autonomous Mode
  • Automatic Context Compaction
  • Retained Subagent
  • Agent-to-Agent Messaging

구조적으로 다음과 같은 실행이 가능하다.

Goal
 │
 ├── Task
 │    └── Subagent
 │
 ├── Heartbeat
 │
 ├── Scheduled Re-entry
 │
 ├── Context Compaction
 │
 └── Refinement

단일 Conversation 종료와 Agent 실행 종료가 반드시 일치하지 않는 구조이다.


5. Hermes Agent

5.1 주요 목적

Hermes Agent는 Coding Agent에 한정되지 않는 범용 Autonomous Agent에 가깝다.

현재 다음 기능들을 제공한다.

  • Persistent Memory
  • Skill
  • Subagent
  • Tool Calling
  • MCP
  • Terminal
  • Browser
  • Messaging Gateway
  • Scheduled Task
  • Voice
  • 사용자 Profile
  • Cross-session Recall

CLI뿐 아니라 Telegram, Discord, Slack, WhatsApp, Teams 등 다양한 Messaging Interface를 지원한다.


6. Hermes의 Self-Improvement

Hermes는 공식적으로 Closed Learning Loop를 제공한다.

Conversation / Task
        │
        ▼
     Experience
        │
        ├── Memory Review
        │
        └── Skill Review
              │
              ▼
       Persistent State
        ├── Memory
        └── Skill

Hermes 문서에서 Self-Improvement의 주요 구성은 다음과 같이 정의된다.

  • Agent-curated Memory
  • Periodic Memory Nudge
  • Autonomous Skill Creation
  • Skill Self-improvement
  • Cross-session Recall
  • User Modeling

즉 Hermes의 자기개선은 주로 다음 두 영역에 집중되어 있다.

What should I remember?
        → Memory

How should I perform this task?
        → Skill

7. Hermes Memory

Hermes는 지속 정보를 여러 역할로 구분하여 관리한다.

파일 역할
USER.md 사용자 정보 및 선호
MEMORY.md 환경 정보, 결정, 경험, 학습 결과
SOUL.md Agent의 Identity 및 동작 스타일
AGENTS.md Project 단위 지침

USER.md 및 MEMORY.md는 Agent가 Memory Tool을 통해 직접 관리할 수 있다.

Memory는 Session 간 유지되며 Session 시작 시 Agent Context에 반영된다.


8. Hermes Skill

Hermes에서는 Skill을 Procedural Memory로 사용한다.

일반적인 구조는 다음과 같다.

skill-name/
├── SKILL.md
├── scripts/
└── references/

SKILL.md에 작업 절차 및 사용 조건을 기록하고 필요할 경우 Script 또는 Reference를 포함할 수 있다.

따라서 Hermes의 Skill은 기본적으로 다음 형태이다.

Procedure
+
Instructions
+
Optional Scripts

Hermes는 작업 경험을 기반으로 Skill을 생성할 수 있고 기존 Skill을 사용하면서 수정할 수도 있다.


9. 자기개선 방식 비교

9.1 개선 대상

항목 Prime Agent Hermes Agent
Memory 지원 지원
사용자 Profile 제한적 지원
Skill 지원 지원
Skill 자동 생성 Skill Creator 사용 지원
Skill 자동 개선 지원 지원
Supplemental Prompt 지원 제한적
Base System Prompt 변경 미지원 미지원
Subagent Specification 개선 지원 제한적
Agent 실행 구조 개선 지원 범위에 포함 Skill 중심
Cross-session Memory 지원 지원
Rollback Refinement Snapshot Workspace Checkpoint 등
Model Weight Training 미지원 미지원

두 프로젝트에서 사용하는 "Self-Improvement"는 Model Weight를 직접 업데이트하는 Online Training을 의미하지 않는다.

현재 공개된 기능은 주로 Agent 외부 상태와 실행 방법을 변경하는 Inference-time Adaptation에 해당한다.


10. 자기개선 범위 차이

Hermes의 Learning Loop는 다음과 같이 정리할 수 있다.

Experience
   │
   ├── What was learned?
   │        ↓
   │      Memory
   │
   └── How was the task solved?
            ↓
           Skill

Prime Agent의 Continual Harness는 범위를 다음과 같이 확장한다.

Experience
   │
   ├── What was learned?
   │        → Memory
   │
   ├── How should the task be performed?
   │        → Skill
   │
   ├── What instructions should be retained?
   │        → Prompt
   │
   └── What Agent structure should be used?
            → Subagent Spec

따라서 자기개선 상태의 범위는 다음과 같이 구분할 수 있다.

Hermes
Memory + Skill

Prime Agent
Memory + Skill + Prompt + Subagent Specification

11. Tool 실행 구조 비교

Hermes

Hermes는 일반적인 Agent Tool Calling 구조를 사용한다.

LLM
 │
 ├── Terminal Tool
 ├── File Tool
 ├── Browser Tool
 ├── Memory Tool
 ├── MCP Tool
 └── Delegate Tool

Tool은 각각 Agent에게 Function 형태로 제공된다.


Prime Agent

Prime Agent는 Persistent IPython을 중심으로 Tool 사용을 프로그램화한다.

LLM
 │
 └── IPython
       │
       ├── File
       ├── Shell
       ├── MCP
       ├── Python Library
       └── Subagent

이에 따라 여러 Tool을 사용하는 절차 자체를 Python 프로그램으로 구성할 수 있다.

results = await asyncio.gather(
    rlm(task_a),
    rlm(task_b),
    rlm(task_c),
)

result = aggregate(results)

이 방식은 Tool Selection뿐 아니라 Loop, Parallel Execution, Filtering, Aggregation 등의 Control Flow를 모델이 직접 구성할 수 있도록 한다.


12. 장점 비교

12.1 Prime Agent

Programmatic Orchestration

Tool 호출과 Agent 호출을 Python 프로그램 형태로 구성할 수 있다.

복잡한 Tool Chain을 다음과 같이 모델 내부에서 구성할 수 있다.

Search
   ↓
Filter
   ↓
Parallel Subagents
   ↓
Verification
   ↓
Aggregation

Context 관리

모든 Tool Schema를 지속적으로 Prompt Context에 유지하는 구조에 대한 의존도가 낮다.

Subagent

Recursive Subagent가 기본 실행 모델에 포함되어 있으며 Agent 간 직접 통신도 가능하다.

Continual Harness

Memory 이외에도 Prompt, Skill, Subagent Specification을 지속 상태로 관리한다.

Long-running Task

Daemon, Goal, Heartbeat, Schedule, Session Reattach 등을 기본 기능으로 제공한다.


12.2 Hermes Agent

Closed Learning Loop

Memory 및 Skill 생성과 개선을 위한 Learning Loop가 제품 기능으로 통합되어 있다.

Persistent User Context

사용자 Profile, Memory, Session Search를 통해 사용자 단위 장기 Context 관리 기능을 제공한다.

Skill Ecosystem

SKILL.md 기반 Skill 체계를 사용하며 기존 Tool과 Script를 절차화하여 재사용할 수 있다.

운영 Interface

CLI 외에 Messaging, Voice, Scheduled Task 등 다양한 실행 Interface를 제공한다.

Runtime 선택

Local, Docker, SSH 및 Remote Sandbox 등 다양한 Terminal Backend를 지원한다.


13. 제약사항 비교

13.1 Prime Agent

Security Isolation

Prime Agent는 모델이 생성한 Python 및 Project Command를 사용자 권한으로 실행한다.

공식 README에서는 Worker와 Kernel Process Separation이 Lifecycle Isolation 및 Recovery 목적이며 Security Sandbox가 아니라고 명시하고 있다.

따라서 신뢰할 수 없는 Code 또는 Instruction을 실행하는 경우 별도의 Sandbox가 필요하다.

Model Capability 의존성

Programmatic Tool Calling과 Recursive Agent 구성을 모델이 직접 수행하므로 모델의 Code Generation 및 Planning 능력에 영향을 받는다.

Prime Intellect 역시 현재 Frontier Model이 새로운 Harness 구조의 전체 기능을 충분히 활용하지 못하고 있으며 향후 Harness를 고려한 Model Training 가능성을 제시하고 있다.

프로젝트 성숙도

Prime Agent는 2026년 8월 5일 처음 공개되었으며, 관련 기술 보고서는 2026년 8월 24일 공개되었다. 따라서 공개 프로젝트 기준 운영 사례 및 장기간의 안정성 검증 데이터는 아직 제한적이다.


13.2 Hermes Agent

Tool Context

일반적인 Tool Calling 구조를 사용하므로 Tool 수가 증가할 경우 Tool Definition과 Context 관리가 필요하다.

Skill 중심 개선

Self-Improvement의 중심 대상은 Memory 및 Skill이다.

Agent Prompt나 Subagent Topology 전체를 하나의 지속적인 개선 상태로 관리하는 구조는 Prime Agent의 Continual Harness보다 제한적이다.

자동 학습 결과의 품질 관리

Agent가 생성하거나 수정한 Memory 및 Skill은 이후 작업에 재사용되므로 잘못된 경험이 지속 상태에 반영될 가능성이 존재한다.

Hermes는 Memory Write Approval 등의 제어 기능을 제공한다.


14. 보안 구조 비교

Hermes는 다음과 같은 Defense-in-Depth 보안 기능을 제공한다.

  1. 사용자 Authorization
  2. Dangerous Command Approval
  3. File Write Safety
  4. Container Isolation
  5. MCP Credential Filtering
  6. Context File Prompt Injection Scan
  7. Cross-session Isolation
  8. Input Sanitization

특히 Docker 등 Container Backend를 실행 경계로 사용할 수 있다.

Prime Agent는 현재 자체 Worker/Kernel 분리를 Security Boundary로 간주하지 않으며 별도의 External Sandbox 사용을 권장한다.

항목 Prime Agent Hermes Agent
Process Isolation 지원 지원
Security Sandbox 기본 제공 아님 Container Backend 지원
Dangerous Command Approval 제한적 지원
File Write Guard 제한적 지원
MCP Credential Isolation 별도 구성 필요 지원
Prompt Injection 대응 별도 구성 필요 지원
외부 Sandbox 권장 선택 가능

15. 기능 성격 비교

구분 Prime Agent Hermes Agent
주요 성격 Coding / Research Harness General Autonomous Agent
핵심 추상화 RLM + Continual Harness Memory + Skill Learning Loop
Tool 실행 Programmatic Function Calling 중심
Subagent Recursive Agent Delegation
Memory Harness State 일부 주요 기능
Skill Executable Python 중심 SKILL.md 중심
Prompt 개선 Supplemental Prompt 지원 제한적
Subagent Spec 개선 지원 제한적
사용자 Profile 제한적 주요 기능
Long-running Agent 주요 기능 지원
Messaging 중심 기능 아님 주요 기능
보안 Isolation 외부 Sandbox 필요 Container Backend 제공
프로젝트 공개 2026-08 기존 프로젝트
주요 적용 영역 Coding, Research, Evaluation Personal Agent, Automation, Coding

16. Self-Improvement 개념 비교

두 시스템의 차이는 다음과 같이 요약할 수 있다.

Hermes Agent

Experience
    ↓
Learning Loop
    ↓
┌──────────────┐
│ Memory       │
│ Skill        │
└──────────────┘
    ↓
Future Session

Hermes의 주요 목적은 과거 경험을 Memory 및 Procedural Skill로 축적하여 이후 작업에 재사용하는 것이다.

Prime Agent

Experience
    ↓
Refinement
    ↓
┌─────────────────────┐
│ Prompt              │
│ Memory              │
│ Skill               │
│ Subagent Spec       │
└─────────────────────┘
    ↓
Future Execution

Prime Agent에서는 개선 범위를 Memory 및 Skill 외에 Prompt와 Agent 구성 정보까지 확장한다.


17. 주요 차이 요약

구분 Hermes Agent Prime Agent
자기개선의 중심 경험 학습 Harness 개선
주요 지속 상태 Memory, Skill Prompt, Memory, Skill, Subagent Spec
개선 자동화 Background Learning Loop 중심 /refine 및 Harness Refinement 중심
Skill 형태 절차 문서 + 선택적 Script Python Executable Skill
Agent 구성 개선 제한적 Subagent Spec 관리
Orchestration Tool Call 및 Delegation Python Program + Recursive Agent
장기 사용자 Memory 강함 지원
장기 실행 지원 핵심 설계 요소
안전 실행 Container 및 Approval 기능 외부 Sandbox 의존
주요 사용 목적 범용 개인/업무 Agent Coding/Research Harness

18. 결론

Prime Agent와 Hermes Agent는 모두 Self-Improving Agent를 표방하지만 개선 대상에는 차이가 있다.

Hermes Agent는 Memory와 Skill을 중심으로 경험을 지속적으로 축적하는 구조를 제공한다. 사용자의 정보, 작업 환경, 과거 Session 및 작업 절차를 장기적으로 저장하고 재사용하는 Closed Learning Loop가 주요 특징이다.

Prime Agent는 Agent Harness 자체를 지속 상태로 관리하는 구조를 사용한다. Continual Harness를 통해 Memory뿐 아니라 Supplemental Prompt, Skill, Subagent Specification까지 관리하고 Refinement 대상으로 포함한다.

또한 Prime Agent는 Persistent IPython과 Recursive Language Model을 사용하여 Tool 및 Subagent 실행 자체를 Programmatic Orchestration 형태로 구성한다.

따라서 두 시스템의 자기개선 범위는 다음과 같이 정리할 수 있다.

Hermes Agent
Experience
→ Memory / Skill
→ 이후 작업에 재사용

Prime Agent
Experience
→ Prompt / Memory / Skill / Subagent Spec
→ Agent Harness에 반영
→ 이후 실행 방식에 재사용

현재 공개 구현 기준으로 Hermes Agent는 Memory, Skill, Messaging, Security 및 다양한 Runtime을 포함하는 범용 Agent Platform의 성격이 강하며, Prime Agent는 RLM과 Continual Harness를 중심으로 Long-running Coding 및 Research Agent의 실행 구조를 확장하는 성격이 강하다.

두 프로젝트 모두 현재의 자기개선을 Model Weight 자체의 Online Learning으로 구현하지는 않으며, 주로 Inference-time 상태와 Agent 실행 방법의 지속적인 변경 및 재사용을 통해 Self-Improvement를 구현한다.


참고 자료

  • Prime Agent GitHub: PrimeIntellect-ai/prime-agent
  • Prime Agent 공개 문서, 2026-08-05
  • Prime Agent: A Self-Improving RLM Harness, 2026-08-24
  • Hermes Agent 공식 Documentation
  • Hermes Agent GitHub
  • Hermes Skills Documentation
  • Hermes Security Documentation
댓글
최근에 올라온 글
최근에 달린 댓글
«   2026/08   »
1
2 3 4 5 6 7 8
9 10 11 12 13 14 15
16 17 18 19 20 21 22
23 24 25 26 27 28 29
30 31
Total
Today
Yesterday