What each aisle in the frameworks grocery store offers—and how the 9 main architecture patterns work under the hood.
Each aisle groups a type of capability you can add to your assistant. Think of them as sections in a grocery store — you choose the ingredients that make sense for your project.
Defines who your AI is—its tone, values, way of speaking, and how it adapts to different contexts. Without this, you have a generic chatbot. With it, you have an assistant with its own voice.
Layers of protection between your AI and the outside world—from tool sandboxing to prompt injection attack detection. Essential in any environment that accesses sensitive data.
The channel is your AI's "point of entry." It determines which app the user will interact with it in. Each channel has distinct characteristics — choose where your audience already is.
The difference between an AI that forgets everything when it restarts and one that has known you for months. The memory system you choose determines the cost, accuracy, and long-term recall capability.
Turns your AI from reactive to proactive. Instead of waiting for you to ask something, it runs tasks at set times, monitors events, and works in the background.
How your AI connects to model providers, external services, and standard protocols. This is where decisions about AI model cost, privacy, and scalability are made.
The longest corridor — these are the concrete capabilities your AI can perform today, without building them from scratch. It covers productivity, communication, smart home, development, and much more.
The structural patterns that define how your assistant’s components connect. Your choice here affects maintainability, scalability, and ease of extension.
How each framework organizes its components internally. These are the blueprints you'd follow when building with each approach.
OpenClaw uses a central WebSocket server (Gateway) as the single entry point for all channels. Messages from WhatsApp, Telegram, Slack, iMessage, and voice pass through the same hub before reaching the agent. This provides centralized session control, routing between multiple agents, and unified support for visual canvas and voice.
Each subsystem (AI provider, channel, memory) is defined by a trait — a contract that says "any implementation must have these methods." The config.toml file decides which implementations to use at runtime. You can switch the AI provider from Claude to Ollama, or the channel from Telegram to WhatsApp, without touching the agent's code. The result is a memory footprint below 5MB and boot in 10ms.
NanoClaw solves a specific WhatsApp problem for businesses: how do you ensure that the context of a work group never leaks into the family group? The answer is radical — each conversation runs in its own isolated container, with its own CLAUDE.md file and separate filesystem. The host manages the WhatsApp connection and dispatches messages to the right container.
TinyClaw treats each agent as a specialized team collaborator. Messages enter a file-based queue (no database), and each agent picks them up from its own inbox. One agent can mention another (@reviewer, @writer) to pass the work along. Parallel processing: while @coder writes code, @reviewer can be analyzing another PR.
IronClaw starts from the principle that no input is trustworthy — not user messages, tool results, or LLM outputs. Each stage of the pipeline has its own security layer: prompt injection detection on input, a WASM sandbox for tool isolation, and a secret leak detector on output. The Job Scheduler allows up to 5 parallel executions with rate limiting.
Agent Zero is built around a continuous improvement loop: before solving a problem, it checks memory ("Have I solved this before?"). Once it solves it, it saves the solution for future use. For complex tasks, it delegates to specialized sub-agents for research or code. Over time, it gets more efficient at the tasks you use most — and can rewrite its own behavior rules in the middle of a conversation.
Instead of the single-agent loop (Telegram→loop→tool→memory), ClaudeClaw runs N specialist agents as independent bots on the SAME machine and DB. They delegate to each other (@agente:), meet in real time (text/voice War Room), and publish every action to a shared event bus (hive_mind table) that serves as both team memory AND visualization—a 3D anatomical brain. Alongside this is an operational containment plan: kill switches with hot reload + default-deny tool policy + exfiltration guard form a runtime security mesh over full-access execution.
Hermes is a perceive→act→reflect→save loop. A single gateway normalizes any channel (CLI, Discord, Slack, WhatsApp, HTTP, cron) into a Message object. The reflect step uses a low-cost model to curate memory (local SQLite FTS5/BM25, in 3 layers) and detect repeated workflows — those patterns become auto-generated skills (SKILL.md in the agentskills.io format) that the agent reuses. Everything is model-agnostic (role-based routing) and proactive via heartbeat/cron. It is the counterpoint to frameworks that statically separate channel/loop/memory: here, the agent improves its own memory and skills through use.
AIOS embeds the LLM as the core of a kernel and treats each LLM agent as an operating system PROCESS. Multiple agents coexist, sharing models, memory, and tools; the Scheduler (FIFO/Round-Robin) arbitrates who calls the LLM, the Context Manager takes snapshots and restores them to pause and resume agents, and everything flows through auditable syscalls. The agent communicates with the kernel through the Cerebrum SDK (separate repo), in Local or Remote mode. Frameworks such as LangChain, AutoGen, and CrewAI run ON TOP OF it via an adapter.