17 — Comparisons

What Was Researched

Cross-cutting comparisons across all studied codebases, identifying common patterns, divergent approaches, and relative strengths/weaknesses.

Which Sources Were Used

All 10 local codebases + model landscape research from directory 13.

Key Findings

Master Comparison: Agent Frameworks

Feature Hermes OpenClaw Codex Pi LangGraph
Language Python TypeScript Rust TypeScript Python
License MIT MIT Apache-2.0 MIT MIT
Architecture Monolithic + plugins Gateway-centric Modular crates Minimal monorepo Graph state machine
Agent loop While-loop (sync) While-loop While-loop While-loop Graph traversal
Model-agnostic ✅ 200+ via OpenRouter ✅ Multi-provider ⚠️ OpenAI-focused ✅ Multi-provider ✅ Via LangChain
Tools 87+ files First-class Minimal core + MCP Minimal Via LangChain tools
MCP ✅ Client + Server ✅ Client ✅ First-class ❌ ⚠️ Via integration
Skills ✅ Autonomous ✅ Manual ❌ ❌ ❌
Memory ✅ 8 providers ✅ Files ❌ ❌ ✅ Checkpoints
Subagents ✅ Advanced ✅ Session-based ❌ ❌ ✅ Subgraphs
Voice ✅ Full pipeline ✅ Wake + Talk ❌ ❌ ❌
Channels 20+ platforms 26+ platforms CLI only CLI only N/A (library)
Sandboxing ✅ 7 backends ✅ Docker/SSH ✅ OS-native ✅ External N/A
AGENTS.md ✅ 71KB ✅ ✅ Creator ✅ ⚠️
Cron/automation ✅ ✅ + webhooks ❌ ❌ ❌
Desktop app ✅ Electron ✅ Win/Mac/iOS/Android ❌ ❌ N/A
Test suite ~1,655 test files CI N/A vitest pytest
Community Nous Research 200+ contributors OpenAI Small team LangChain Inc

SDK / Proxy Comparison

Feature LiteLLM OpenRouter SDK LangChain Open Responses
Language Python TypeScript Python Go
Approach SDK + Proxy SDK SDK + integrations API proxy
Providers 100+ 200+ (via OR) 700+ integrations Any OpenAI-compat
Wire format OpenAI-compat OpenAI-compat Custom + adapters OpenAI Responses
Provider translation BaseConfig pattern Speakeasy-gen Provider packages Go structs
Streaming ✅ ✅ ReusableStream ✅ ✅
Caching ✅ Redis N/A ⚠️ ✅ Redis
Self-hostable ✅ N/A (client SDK) N/A ✅
Production scale ✅ 8ms P95 ✅ ✅ ⚠️ Single monolith

Model Tier Comparison (June 2026 pricing)

Tier Best Value Best Quality Best for Coding
Frontier Nemotron 3 Ultra ($0.50/M in) Claude Fable 5 ($10/M in) Kimi K2.7 Code ($0.68/M in)
Flash Gemini 3.5 Flash GPT-5.4 Mini DeepSeek V4 Flash
Nano GPT-5.4 Nano — —
Embedding Text Embedding 3 Small ($0.02/M) Text Embedding 3 Large ($0.13/M) —
Reranking Rerank 4 Fast Rerank 4 Pro —

Architecture Pattern Comparison

Pattern Proponent Tradeoffs
Monolithic agent Hermes Simple deploy, hard to scale, huge files
Gateway-centric OpenClaw Scalable, distributed, complex setup
Modular crates Codex Maximum modularity, Rust learning curve
Minimal monorepo Pi Clean separation, limited features
Graph state machine LangGraph Best debugging, academic overhead
SDK + Proxy LiteLLM Flexible, dual deployment, two codebases

What Is Confirmed

  1. No single framework has it all — Hermes leads in features, OpenClaw in channels, Codex in security, Pi in simplicity, LangGraph in debuggability
  2. Python + TypeScript is the winning stack — Python for agent logic, TypeScript for gateway/frontend
  3. OpenAI-compatible wire format is the universal choice
  4. AGENTS.md + SKILL.md are the emerging file conventions
  5. MCP is the extensibility standard — all major frameworks support it
  6. 20x pricing spread in models means intelligent routing is critical

How This Applies to Building a Modern Model-Agnostic Agent Harness

The ideal harness would combine:

  • Hermes's learning loop (skills, memory, autonomous improvement)
  • OpenClaw's gateway architecture (multi-channel, distributed)
  • Codex's security model (sandboxing, context discipline)
  • Pi's clean separation (AI → Agent → Application layers)
  • LangGraph's state management (checkpointing, interrupts)
  • OpenAI-compat model access (Ollama local, OpenRouter hosted, or optional LiteLLM proxy — not all three)