Hermes/HAI vs Mainstream Agent Frameworks (2025–2026)
Quelle: Subagent Framework-Vergleich
Samuel's stack is best understood as a personal AI operating system, not a reusable framework. It combines Hermes profiles (50+ role-specialized agents), SOUL.md identity contracts, skills as orchestration scripts, MCP bridges to personal data, and HAI as the human-facing decision/execution layer. The "agentische Sanduhr" model splits intake (top pyramid) → human decision surface (middle) → execution (bottom pyramid).
Comparison Table
| Dimension | Samuel Hermes/HAI | LangGraph | CrewAI | AutoGen | OpenAI Agents SDK |
|---|---|---|---|---|---|
| Architecture | Profile-per-role OS: global SOUL.md + per-profile SOUL.md, config.yaml, skills, hooks. HAI = planner → executor → validator chain. Intake-router, KG, Codex/Cursor as worker lanes. "Two pyramids" sanduhr model. |
Graph: typed State + Nodes + Edges. Pregel-style supersteps. Subgraphs for multi-agent. |
Dual layer: Crews (autonomous role teams) + Flows (event-driven orchestration with @start/@listen/@router). Standalone since v1.0 (no LangChain). |
Layered actor model (Core → AgentChat → Extensions). Async message-passing agents. Maintenance mode — successor is Microsoft Agent Framework (typed graph Workflow). |
Minimal primitives: agent loop, handoffs, Agent.as_tool(), sessions. Thin runtime over Responses API. |
| State | File artifacts in Projek-Managment/: PROJECT_STATE.md, BRIEFING.md, NEXT_STEP.md, execution/validation reports. Lossless archives (archive/next_steps/, NEXT_STEP_HISTORY.md). Profile memory + KG MCP. Checkpoints disabled in HAI config. |
First-class typed state with reducers (operator.add for messages). Checkpointers (Memory/SQLite/Postgres) per thread_id. Time-travel, replay, resume. |
Flow state: dict or Pydantic model. @persist decorator for snapshots. Crew memory hierarchy. |
Event-sourced message history in runtime. MAF adds graph-based workflow state with superstep checkpointing. | RunState serializable to JSON. Sessions, conversationId, previousResponseId strategies. Durable pause/resume boundary. |
| Tool use | Per-profile toolsets (terminal, file, skills, MCP). MCP: Gemini, Knowledge Graph (kopf), Hermes bridge, gdrive, plaud, hai-intake. Tirith security gate. Skills gate which tools run when. External lanes: OpenCode, Cursor Composer, Codex. |
Native tool nodes in graph. MCP via LangChain ecosystem. interrupt() inside tool functions for approval. |
Per-agent tools + LiteLLM model routing. MCP/A2A support in recent versions. Tool abstractions owned by CrewAI. | Tool execution via agent chat + extensions. MAF adds hosted tools, middleware, conversation threads. | Built-in tool loop. needsApproval on tools. Hosted MCP tools. Guardrails. Sandbox agents for file workspaces. |
| Human-in-the-loop | Constitutional, not mechanical: Samuel's autonomy > agent judgment. Explicit gates for commit/push/deploy/destructive/auth. HAI compresses decisions (recommended default + max 2 alternatives). hai-minimize reduces cognitive load. Feierabend stop. No programmatic interrupt/resume API. |
interrupt(value) + Command(resume=...). Checkpointer mandatory. Approve/reject/edit tool calls. interrupt_before/after compile-time breakpoints. Production-grade pause for days/months. |
Human feedback gates in Flows. @persist for pause/resume. Role-based delegation with manager agent in hierarchical process. |
Observation/control affordances in event runtime. MAF inherits enterprise HITL patterns from Semantic Kernel. | Native approval flow: RunResult.interruptions → state.approve()/reject() → Runner.run(agent, state). Works across handoffs and nested as_tool(). |
| Samuel does better | • Deep personal context (ADHS profile, uni, startup, voice intake → projects) • Role hardening via SOUL boundaries, not just prompts • Decision compression (HAI minimize, default recommendation) • Audit trail (artifact chain, capability ledger, lossless archives) • Multi-channel OS (Telegram/Discord/WhatsApp/Slack + TTS/STT) • Meta-governance (meta-new-agent reviews new profiles) • Life-domain specialization (50+ profiles: uni, calendar, intake, HAI) |
— | — | — | — |
| Samuel does worse | • No typed graph or formal state machine • No checkpoint/resume API (manual file state) • Operational complexity (profile sprawl, parallel-session overload — documented bottleneck) • Not portable/distributable • Weaker testability & reproducibility • Single-maintainer burden • No enterprise observability/tracing out of the box | • Generic — no personal context layer • Steeper graph design learning curve • Less "human decision UX" by default | • Role metaphor can hide scope creep • Production patterns still require Flow wrapping • Less personal-OS integration | • Effectively frozen — new work → MAF • Distributed runtime still experimental • Conversation-centric, less artifact-oriented | • OpenAI-centric • Less multi-profile identity model • No built-in intake/life-OS layer • Sessions ≠ project artifact governance |
Bottom Line
Samuel's stack is stronger as a personal cognitive prosthesis — context, boundaries, decision load, and auditability are first-class. Frameworks are stronger as engineering substrates — typed state, interrupt/resume, scale, and ecosystem.
The documented HAI bottleneck (execute phase + parallel session overload) is exactly where LangGraph-style checkpointing and Flow-level orchestration would help — but migrating SOUL/profile/artifact semantics into a framework would lose much of what makes the personal OS work.