Reference Guide · Updated April 2026

The AI Agent
Landscape

By Craig Hepburn

Section I — Personal & Always-On Agents Section II — Coding Agents · Platforms · Frameworks

Section I

Personal & Always-On Agents

Self-hosted agents that run 24/7, connect to your messaging apps, and act on your behalf. This category was created by OpenClaw in late 2025 and has spawned an entire ecosystem in under six months.

The originals — defined the category

Category founder

openclaw.ai ↗

Peter Steinberger → OpenAI / Foundation

The original viral personal agent. Gateway architecture routes all messages through a central controller. 5,700+ community skills on ClawHub. Model-agnostic, 50+ integrations. Now under independent foundation with OpenAI backing after Steinberger joined OpenAI Feb 2026.

345K+ starsTypeScript500K lines
Self-improving

github/NousResearch/hermes-agent ↗

Nous Research

Agent-loop architecture. Persistent memory, autonomous skill creation, self-improving learning loop, user modelling via Honcho. Positioned between Claude Code CLI and OpenClaw. Built-in hermes claw migrate command for OpenClaw users.

10K+ starsPythonFeb 2026

The security response — built because OpenClaw's CVEs were alarming

Security-first

github/qwibitai/nanoclaw ↗

Gavriel Cohen / Qwibit

~700 lines of TypeScript — auditable in 8 minutes. Every agent runs in an isolated Linux container that self-destructs after each message. Built on Claude Agent SDK. Container-per-chat-group isolation prevents data leakage between contexts.

21K+ starsTypeScriptJan 2026
Enterprise security

ironclaw.com ↗

NEAR AI

Rebuilt from scratch in Rust. WASM sandboxing with cryptographic capability tokens per skill. TEE-backed execution, encrypted credential vaults, zero telemetry. PostgreSQL with AES-256-GCM. The only personal agent suitable for regulated industries.

RustEnterpriseEarly 2026
NVIDIA-backed

github/NVIDIA/NemoClaw ↗

NVIDIA

Not a standalone agent — a security wrapper over OpenClaw. Adds OpenShell kernel-level sandbox via Linux security modules. Policy-as-YAML. Ships with NVIDIA Nemotron open models. Announced by Jensen Huang at GTC March 2026.

TS plugin + PythonPreview Mar 2026

The minimalists — lighter, hackable, hardware-friendly

Minimal Rust

zeroclawlabs.ai ↗

Community

Runtime rewritten in Rust. 3.4MB single binary. Zero cloud dependencies, entirely local. Ultra-low memory footprint for users who want complete data sovereignty.

Rust3.4MB binary
Minimal Python

github — community ↗

Community

4,000 lines of readable Python. Runs on Raspberry Pi. Supports local models via Ollama. No plugin system, no config sprawl. Built for developers who want to understand every line of their agent.

Python4K linesPi-ready
Ultra-minimal

github/topics/picoclaw ↗

Community

Runs on $10 boards. Under 10MB RAM. Designed for constrained hardware — Raspberry Pi Zero, embedded systems. Single-purpose, maximum portability.

Embedded<10MB RAM

Multi-agent and enterprise variants

Multi-agent OS

github/agentscope-ai/hiclaw ↗

Community

Collaborative multi-agent OS built on Matrix rooms. Transparent, human-in-the-loop task coordination. Supports OpenClaw, NanoClaw, and ZeroClaw agents as workers. MinIO shared filesystem for inter-agent exchange.

Multi-agentMatrix protocol
Desktop agent

eigent.ai ↗

Eigent

Desktop multi-agent workforce. Connects to your context, controls browser and desktop apps to automate real work. Visual interface rather than messaging-native. Non-developer accessible.

DesktopCommercial
Dev channels

claude.ai/download ↗

Anthropic

Not a full agent framework. Lets developers control local Claude Code sessions from Telegram and Discord. Specifically for developer workflows, not general personal assistant use. Launched March 2026.

Dev workflow onlyMar 2026

Head-to-head comparison

Yes — full native support ~ — partial or conditional No — not supported N/A — not applicable
OpenClaw Hermes NanoClaw IronClaw NemoClaw ZeroClaw Nanobot
Architecture
Core designGateway / controllerAgent loop / learningContainer-per-agentWASM + RustWrapper over OpenClawSingle binary / RustSimple loop / Python
LanguageTypeScriptPythonTypeScriptRustTS + PythonRustPython
Codebase size500K lines~8,900 lines core~700 linesRust, auditableWrapper3.4MB binary4,000 lines
Self-improving / learnsNoYesNoNoNoNoNo
Persistent memoryContext fileYes, nativePer-containerEncrypted DBVia OpenClawPartialBasic
Messaging channels
WhatsAppYesYesYesVia pluginsVia OpenClawPartialVia config
TelegramYesYesYesYesVia OpenClawYesYes
Slack / DiscordYesYesYesYesVia OpenClawPartialYes
SignalYesYesPartialNoVia OpenClawNoNo
iMessageYesNoNoNoVia OpenClawNoNo
Security
Container / sandbox isolationNo6 backendsYes, per-agentWASM + TEEKernel sandboxProcess onlyNo
Prompt injection scanningPatched CVEsYes v0.7+Via isolationiron-verifyYesBasicNo
Known critical CVEs3 criticalNoneNoneNoneNoneNoneNone
Suitable for regulated industriesNoWith hardeningWith hardeningYesYesPartialNo
Ecosystem & cost
Skills / plugin marketplaceClawHub 5,700+agentskills.ioFork-basedNoVia OpenClawNoNo
MCP supportYesYesVia extensionVia pluginsVia OpenClawNoNo
Local model support (Ollama)YesYesVia Claude SDKYesNemotron bundledYesYes
Managed hosting availableYes ($59/mo)Partners (~$14/mo)Self-host onlyEnterpriseN/ASelf-host onlySelf-host only
Runs on low-power hardware1GB+ RAMVPS min.YesModerateHost onlyYesRaspberry Pi

The lineage in one sentence: OpenClaw proved the category existed. NanoClaw and IronClaw were built because OpenClaw's security was dangerous. Hermes was built because OpenClaw didn't learn. ZeroClaw and Nanobot were built because OpenClaw was too heavy. NemoClaw was built because NVIDIA saw the enterprise opportunity. HiClaw was built because nobody had coordinated multiple agents yet.

Section II

Coding Agents, Managed Platforms
& Developer Frameworks

The broader agent landscape: terminal-native coding tools, hosted infrastructure for production deployment, big tech model-native agent systems, and the developer frameworks used to build custom architectures.

Coding agents — terminal-native developer tools

Open source · engine

pi.dev / shittycodingagent.ai ↗

Mario Zechner

Minimal terminal coding harness. Only 4 core tools: read, write, edit, bash. Maximally extensible via TypeScript extensions. The engine that powers OpenClaw. Model-agnostic across 15+ providers. Mid-session model switching.

14K+ starsTypeScriptMIT
Managed

claude.ai/download ↗

Anthropic

Batteries-included CLI. Deep codebase context, git workflows, IDE integration (VS Code, JetBrains, Xcode). Included in Pro/Max plans. SOC2 Type II. Foundation for Managed Agents. Best-in-class for Claude models.

$20/mo planSOC2 Type II
xAI

x.ai/grok ↗

xAI (now SpaceX)

Local-first CLI coding agent. Up to 8 parallel agents on a single project. Air-gap compatible, no code sent to xAI servers. Native X/real-time data advantage. Waitlist as of Q1 2026.

Local-first8 parallel agents
Open source

github/block/goose ↗

Block

Open-source local agent from Block. CLI and desktop, MCP extensibility, real engineering task automation. Model-agnostic. Privacy-first Claude Code alternative.

Open sourceMIT
Open source

aider.chat ↗

Community

Git-native CLI coding agent. Excellent multi-file editing, strong git integration. Simpler than Pi. Best for developers who want AI-assisted commits without setup overhead or extension work.

Open sourcePython
Google

github/google-gemini/gemini-cli ↗

Google

Terminal coding agent with A2A protocol and Vertex AI backend. Native Google Workspace and Search grounding. Strong for GCP workflows. Part of Google ADK ecosystem.

A2A protocolVertex AI

Managed agent platforms — hosted production infrastructure

Managed cloud

platform.claude.com/docs ↗

Anthropic

Hosted harness with durable session logs, sandboxed cloud containers, crash recovery via wake(sessionId), credential vaults where secrets never reach the sandbox, and MCP tool integration. Brain decoupled from hands. The architecture described in Anthropic's engineering blog.

$0.08/hr + tokensSOC2 Type IIBeta Apr 2026
Managed cloud

platform.openai.com/docs/guides/agents ↗

OpenAI

Managed runtime with built-in web search, code interpreter, file search, agent handoffs, and guardrails. Tightly OpenAI-coupled. Lowest setup friction for GPT-based deployments. Production-ready.

Token-basedPython / TS SDK
Google Cloud

cloud.google.com/products/agent-builder ↗

Google

Enterprise agent platform on GCP. A2A protocol, Gemini models, native Google Workspace connectors. Strong for organisations already in GCP. Managed compliance, observability, and data residency.

GCP-nativeA2A protocol
Microsoft

microsoft.com/copilot-studio ↗

Microsoft

No-code/low-code agent builder on M365. Deep Teams, Outlook, Dynamics integration. Now powered by Claude (Cowork partnership). Best for M365 organisations. Enterprise governance built-in.

M365 licencePowered by Claude
No-code

lindy.ai ↗

Lindy AI

No-code personal agent platform. Connects Gmail, Calendar, Slack, CRMs. Automations from natural language. Consumer-friendly entry for non-developers wanting agentic workflows without self-hosting.

Free / paid tiersWeb app
Managed cloud

Perplexity Computer

perplexity.ai/products/computer ↗

Perplexity AI

Multi-model cloud agent orchestrating 19 AI models simultaneously — routing each subtask to the optimal model. Tasks run in isolated Firecracker microVMs and can persist for hours or months. Also ships Personal Computer, a local variant running on Mac Mini for file and app access. Model-agnostic by design: no single provider dominates.

$200/mo Max plan19 modelsFeb 2026

Big tech — model-native and internal agent systems

xAI model-native

x.ai/grok ↗

xAI (now SpaceX)

4-agent council baked into inference at runtime: Grok (Captain), Harper (research), Benjamin (math/code), Lucas (strategy). Not a framework you orchestrate — runs natively on every complex query. 16 agents on high reasoning mode.

4–16 agents nativeGrok 4.20
Meta internal

engineering.fb.com ↗

Meta

Internal agent framework powering Meta's Ranking Engineer Agent. Multi-week autonomous ML experimentation via hibernate-and-wake mechanism. Not publicly available — internal Meta infrastructure only.

Internal onlyNot public
Microsoft research

github/microsoft/autogen ↗

Microsoft Research

Conversation-based multi-agent framework. Agents communicate via structured chat. Strong human-in-the-loop. Azure integration. Open source and enterprise-deployable. v0.4 rewrite introduced graph-based workflows.

Open sourcePython

Developer frameworks — build your own agent architecture

Framework

langchain.com ↗

LangChain Inc.

Largest ecosystem. LangGraph adds graph-based stateful workflows with explicit control flow and state management. Most integrations. LangSmith for observability. High flexibility, high complexity. Best for custom multi-step pipelines.

Open sourcePython / TSLangSmith SaaS
Framework

crewai.com ↗

CrewAI Inc.

Role-based multi-agent orchestration. Define a crew of specialist agents with task delegation. Sequential or hierarchical coordination. Intuitive mental model for team-oriented workflows. Fast prototyping.

Open sourcePythonEnterprise tier
Google

google.github.io/adk-docs ↗

Google

Agent Development Kit. Python-first, multi-agent, A2A protocol. Connects to Vertex AI and Google Workspace data stores. Designed for developers building on Google infrastructure.

PythonA2A protocolApache 2.0
Framework

llamaindex.ai ↗

LlamaIndex

RAG-first agent framework. Best for agents that need to reason over large document corpora. Deep indexing and retrieval primitives. Combines well with CrewAI for document-heavy multi-agent systems.

Open sourcePython
Framework

mastra.ai ↗

Community

TypeScript-first agent framework for web developers. Built-in Studio with traces and token usage. Durable execution, state management, and long-running workflow support. Growing fast in 2026.

Open sourceTypeScript

Platform comparison

Yes — full native support ~ — partial or conditional No — not supported N/A — not applicable
Pi Claude Code Grok Build Claude Managed OpenAI Agents Copilot Studio Grok Multi-Agent LangGraph CrewAI
What it is
Category Coding harnessCoding agentCoding agent Managed infraManaged infraEnterprise platform Model-nativeDev frameworkDev framework
Primary audience Power devsDevs / buildersLocal-first devs Production teamsGPT buildersEnterprise / M365 Research / power usersCustom pipelinesRole-based crews
Deployment
Self-hosted / local YesYesYes NoNoNo N/AYesYes No
Runs 24/7 background NoNoNo YesNoYes N/ANoNo Yes
Model agnostic Yes, 15+Claude onlyGrok only Claude onlyOpenAI primaryClaude + GPT Grok onlyYesYes Yes, 19 models
Data sovereignty YesLocal + AnthropicYes NoNoNo NoYesYes No
Capabilities
Multi-agent orchestration Via extensionYes8–16 parallel PreviewYesYes Yes, nativeYesYes Yes, native
MCP support Via extensionYesNo YesNoNo NoVia pluginsVia plugins No
Real-time web / data Via extensionWeb searchYes + X firehose YesYesYes Yes, native XVia toolsVia tools Yes, Sonar
Code execution (sandboxed) YesYesLocal Cloud sandboxYesLimited N/ALocalLocal
Production architecture
Durable session / crash recovery Tree branchingPartialNo Yes, durable logPartialYes N/ASetup neededNo Yes
Credential vault (secrets from sandbox) NoNoLocal YesYesYes N/ANoNo
Observability / tracing Session treePartialAuditable local YesYesYes N/ALangSmithEnterprise Yes
Cost & access
Free beyond API costs Yes$20/moWaitlist +$0.08/hrToken-basedM365 licence Via subscriptionYesYes $200/mo
Non-developer accessible NoNoNo Console UINoYes YesNoNo Yes
Security / compliance MIT, unauditedSOC2 Type IILocal, unaudited SOC2 Type IISOC2 Type IIEnterprise xAI / SpaceXOpen sourceOpen source Sandbox + human gates

The four-layer model: The model layer (Claude, Grok, Gemini) provides intelligence. The harness layer (Pi, Claude Code, Grok Build) provides the execution loop. The orchestration layer (OpenClaw, Hermes, LangGraph, CrewAI) routes tasks and manages state. The managed infrastructure layer (Claude Managed Agents, OpenAI Agents SDK, Copilot Studio) handles production deployment, security, and crash recovery. Most people conflate these layers. The real competitive advantage lies in the orchestration layer — not the model.