AI Agent Comparisons¶
Summary
Neutral, side-by-side comparisons of the AI agents and agent components covered in this domain. Each page has a feature matrix sourced from the topic pages, a "which one should I pick?" flowchart, and links back to every compared topic. Checked 2026-09-25.
Comparison Notes¶
| Comparison | Products compared | Summary |
|---|---|---|
| OpenClaw vs Hermes Agent vs Claude Code | OpenClaw, Hermes Agent, Claude Code | Three agent philosophies: universal breadth (OpenClaw, 32 channels), self-improvement (Hermes, agent-written skills), and coding depth (Claude Code) |
| Jev vs AnyJev vs LLM Structured Outputs | Jev, AnyJev, structured outputs (LLM Fundamentals) | Three ways to get a typed decision inside an agent loop: hosted calibrated model, self-hosted log-prob readout, or constrained generation |
Landscape¶
The personal-agent space in 2026 organizes around three philosophies:
- Coding depth: agents built for software engineering, such as Claude Code, Codex, and Cursor. They optimize for codebase understanding, multi-file edits, and IDE integration.
- Universal breadth: general-purpose assistants across many messaging channels and devices. OpenClaw (~390k GitHub stars as of 2026-09) is the reference: 32 channels, native apps, and the ClawHub skill registry.
- Self-improvement: agents that learn from their own use. Hermes Agent (Nous Research) creates and patches
SKILL.mdprocedures after non-trivial work and can evolve skills offline with DSPy + GEPA.
Convergence trend
The philosophies are not mutually exclusive. Claude Code added auto memory, subagents, skills, and chat Channels (research preview). OpenClaw can run Claude Code or Codex as its agent runtime and drafts skills through its Skill Workshop. Hermes supports coding over ACP and imports Claude Code and OpenClaw setups. The core design of each tool still decides where it excels.
A second axis sits inside agent loops: the small judgments (route, gate, score) that do not need generation. That is what the Jev comparison covers.
Future Comparisons to Track¶
| Comparison | Rationale |
|---|---|
| LLM Wiki vs RAG vs GraphRAG | Compiled Markdown knowledge vs retrieval at query time; the LLM Wiki topic covers the contrast in prose |
| Spec Kit vs Kiro | Open-source vs commercial spec-driven development, both instances of AI PDLC |
| Codex CLI vs Claude Code | OpenAI's open-source CLI agent vs Anthropic's coding agent |
| Cursor Agent vs Claude Code | IDE-native agent vs multi-surface agent with IDE extensions |
| Devin vs Claude Code | Fully autonomous SWE agent vs human-in-the-loop coding agent |
| OpenClaw vs Ironclaw vs Hermes Agent | Self-hosted general-purpose agents: community fork dynamics and security posture |
These will be written when the vault has topic pages or enough sourced material for each side.