Skip to content

AI Agents

Summary

This domain covers always-on personal AI agents (OpenClaw, Hermes Agent), the decision models and patterns that sit inside agent loops (Jev, LLM Wiki), the AI-native delivery process that puts an agent in every lifecycle phase (AI PDLC), provider data-retention controls for regulated use (Zero Data Retention), and the LLM internals everything else rests on (LLM Fundamentals). Agentic coding assistants such as Claude Code appear here as comparison points, not as topics. As of 2026-09 the two open-source agents are both MIT-licensed and ship several releases a week: OpenClaw 2026.9.6 (foundation-governed, 32 channels) and Hermes Agent 0.21.5 (~28 platforms, self-improving skills). Both carry real CVE histories, so hardening is part of adoption.

Domain Map

How the seven topics relate: LLM Fundamentals underpins everything, the two agent harnesses consume the other topics, and the arrows name the concrete integration or dependency.

flowchart LR
    LF["LLM Fundamentals<br/>transformers, serving,<br/>structured output"]
    OC["OpenClaw<br/>Gateway, 32 channels"]
    HA["Hermes Agent<br/>learning loop, ~28 platforms"]
    JEV["Jev<br/>typed decision model"]
    WIKI["LLM Wiki<br/>agent-maintained Markdown"]
    PDLC["AI PDLC<br/>agent per lifecycle phase"]
    ZDR["Zero Data Retention<br/>provider retention terms"]

    LF --> OC
    LF --> HA
    LF --> JEV
    OC -->|"Decision Models role<br/>(TypeSafe plugin)"| JEV
    HA -->|"hermes claw migrate"| OC
    OC -->|"imports Hermes memory"| HA
    WIKI -.->|"same compounding-memory idea"| HA
    PDLC -.->|"context substrate<br/>before agents"| WIKI
    ZDR -.->|"retention terms for<br/>hosted models"| OC
    ZDR -.-> HA
    ZDR -.-> JEV

Topics

Versions and licenses come from each topic's Key Facts (checked 2026-09-25).

Topic What it is Latest version (date) License
OpenClaw Self-hosted personal assistant: one Gateway connects 32 messaging channels and native apps to any model through swappable agent runtimes; stewarded by the OpenClaw Foundation (501(c)(3), July 2026) 2026.9.6 (2026-09-23); extended-stable 2026.7.35 MIT
Hermes Agent Nous Research's self-improving agent: bounded MEMORY.md/USER.md, agent-written SKILL.md procedures, FTS5 session search, ~28 platforms, seven terminal backends 0.21.5, tag v2026.9.24 (2026-09-24) MIT
Jev TypeSafe AI's "System One" decision model: state plus typed questions in, calibrated choice/score/noul answers out at 70-500 ms and $0.042/MTok input. Direct signups paused since 2026-09-22; Vercel, OpenRouter, and Cloudflare gateways stay open jev-1.13.0 (2026-09-15) Closed weights, hosted API (SDKs MIT)
LLM Wiki Karpathy's pattern: an agent compiles raw sources into an interlinked Markdown wiki (ingest, query, lint) instead of re-deriving answers with RAG Gist as published 2026-04-04; qmd 2.8.3 (2026-08-16) search layer Gist: none stated (checked 2026-09-27); qmd MIT
AI PDLC AI-native product lifecycle pattern: a governed agent per phase over a context substrate, ended by eval gates. Evidence from Freshworks, Shopify/Duolingo mandates, GitHub Spec Kit, and AWS Kiro Pattern (unversioned); Spec Kit 1.0.11 (2026-09-24), Kiro IDE 1.1 (2026-09-14) Spec Kit MIT; Kiro proprietary
Zero Data Retention Contract terms and provider controls that stop LLM providers storing prompts and outputs; per-provider postures and 2026 carve-outs (Anthropic Covered Models, Bedrock retention modes) Policy topic; latest change Bedrock aws_review mode (2026-09-04) n/a
LLM Fundamentals Transformer internals, training pipeline, numeric formats and quantization, serving engines, fine-tuning, RAG, structured output, LLM security Concept topic; e.g. vLLM 0.30.0 (2026-09-22), SGLang 0.5.20 (2026-09-18) Engines Apache-2.0 or MIT

Agent Landscape (2026)

The personal-agent space settled into three philosophies. Each column links to the page that covers it.

Philosophy Representative Focus
Deep coding Claude Code (not a topic here), Codex, Cursor Codebase-aware coding across terminal, IDE, desktop, and web
Universal automation OpenClaw One assistant across 32 messaging channels and paired devices
Self-improving Hermes Agent Persistent memory, autonomous skill creation and patching, offline GEPA skill evolution

Not mutually exclusive

The philosophies are converging: Claude Code now has auto memory and chat channels (research preview), OpenClaw has a Skill Workshop and memory "dreaming", and Hermes supports coding over ACP. Many users run Claude Code alongside OpenClaw or Hermes. Both open-source agents can host Claude Code: OpenClaw through its claude-cli runtime and ACP, Hermes through migration and ACP.

Comparisons

Which Page Do I Need?

The flowchart routes a question to the topic or comparison that answers it; it agrees with the decision flowcharts on the comparison pages.

flowchart TD
    Q{"What are you trying to do?"} -->|"Pick a personal or coding agent"| CMP["OpenClaw vs Hermes vs Claude Code"]
    CMP --> C1{"Main need?"}
    C1 -->|"Chat-app assistant on many channels"| OC["OpenClaw"]
    C1 -->|"Agent that learns your workflows"| HA["Hermes Agent"]
    C1 -->|"Coding in a repo"| CC["Claude Code (external)"]
    Q -->|"Route, gate, or score inside an agent loop"| J["Jev vs AnyJev vs structured outputs"]
    Q -->|"Keep a long-lived knowledge base"| W["LLM Wiki"]
    Q -->|"Restructure product delivery around agents"| P["AI PDLC"]
    Q -->|"Send regulated data to a hosted model"| Z["Zero Data Retention"]
    Q -->|"Understand or self-host the model itself"| L["LLM Fundamentals"]
  • Chat-app assistant with the widest channel reach: OpenClaw, with sandboxing and exec approvals turned on (both are off by default).
  • An agent that gets better at your recurring tasks: Hermes Agent; turn on write_approval if several people talk to it.
  • Deep coding: Claude Code or a similar coding agent; see the comparison.
  • Small, high-volume judgments (routing, guardrails, tool gating): Jev, or AnyJev/structured outputs when you must self-host; see the comparison.
  • Before sending regulated data anywhere: read Zero Data Retention first.

Claw Family Ecosystem

The "Claw" family is the ecosystem of projects built around or extending OpenClaw. All rows were re-checked against the project READMEs on 2026-09-27.

Project Maintainer Purpose
NemoClaw NVIDIA Runs OpenClaw inside NVIDIA OpenShell sandboxes with managed inference and network policy (Apache 2.0, alpha)
claworc Community Multi-instance orchestration dashboard, one container per OpenClaw instance
AutoResearchClaw AIMING Lab 23-stage research pipeline from idea to paper, with human-in-the-loop modes since v0.4.0 (2026-04)
openclaw-termux Community OpenClaw on Android via Flutter + built-in terminal (OpenClaw now also ships an official Android node app)
Paperclip PaperclipAI Orchestration server and UI for teams of agents (OpenClaw, Codex, Claude, Cursor): "if OpenClaw is an employee, Paperclip is the company"
ClawTeam Community Multi-agent swarm coordination with OpenClaw as default agent
clawdbot Legacy Clawdbot name used Nov 2025 - Jan 2026, before the Moltbot and OpenClaw renames; the old repo now serves the OpenClaw README

Supporting Infrastructure

Project Maintainer Purpose
OpenViking ByteDance Context database for AI agents; also a bundled Hermes memory provider
OpenSpace HKU DS Lab Skill evolution layer across OpenClaw, nanobot, Claude Code, Codex, Cursor
Lobster OpenClaw Typed workflow runtime with approval gates (optional @openclaw/lobster plugin)

Sources

Questions

Open

  • Does ClawHub's VirusTotal/ClawScan scanning keep malicious skills out at scale, compared with the 341-824 found before scanning began?
  • How will OpenClaw Foundation governance (donors include OpenAI) keep the roadmap neutral across model labs?
  • Will Hermes Agent's offline DSPy + GEPA self-evolution (Phase 1: skills only) become a standard pattern, and when do Phases 2-5 ship?
  • Does NemoClaw's OpenShell sandboxing (still alpha) scale to enterprise deployments?
  • Can Lobster plus Task Flow compete with LangGraph-style orchestration frameworks?
  • When will Jev direct signups reopen, and will TypeSafe publish rate limits, an SLA, and data-retention terms?

Answered

  • What is the "Claw" family? Projects extending OpenClaw: NemoClaw (NVIDIA sandboxing), claworc (orchestration), AutoResearchClaw (research), openclaw-termux (Android), Paperclip (enterprise), ClawTeam (swarms).
  • Is OpenClaw related to Claude/Anthropic? Only by name history: Clawd/Clawdbot was renamed to Moltbot (2026-01-27) after Anthropic's trademark request, then to OpenClaw (2026-01-30). OpenClaw can drive Claude models and the Claude Code CLI like any other provider or runtime.
  • Who governs OpenClaw after Steinberger joined OpenAI? The OpenClaw Foundation, launched as a US 501(c)(3) in July 2026, employs the core team and signs releases (OpenClaw topic).
  • Does Hermes Agent really have zero CVEs? No. At least ten CVEs were published against it between April and September 2026 (Hermes security).
  • Which agent for coding vs personal automation? Claude Code for deep coding, OpenClaw for multi-channel automation, Hermes Agent for workflows that should improve with use. Many users combine two.