Framework Watch #1: The Agent Framework Landscape in March 2026 — And How Paperclip Works With All of Them

The agent framework space is moving fast. Here’s what matters, what changed, and where Paperclip fits.

If you’re building with AI agents in 2026, you have more options than ever — and more confusion to go with them. Claude Code ships Agent Teams. OpenClaw goes viral. Microsoft retires AutoGen. MCP becomes the USB port of AI. And somewhere in the middle, you’re trying to figure out which framework to actually bet on.

Here’s the thing: you don’t have to pick just one. Paperclip is the orchestration layer that sits above all of them. Think of it as the company that employs the agents, regardless of which runtime they use.

This is the first edition of Framework Watch — our ongoing look at the agent framework landscape, written for builders who want to stay current without reading every changelog.

The Big Picture: Three Layers of the Agent Stack

Before diving into individual frameworks, it helps to see the stack clearly:

Most frameworks live in the Runtime layer. Some span Runtime and Protocol. Paperclip is in the Orchestration layer — which is why it complements rather than competes with everything below it.

Framework by Framework: What’s New, What Matters

Claude Code — The Power User’s Choice

Anthropic’s Claude Code is the most capable agentic coding tool available today. With Opus 4.6 and a 1M token context window, it handles complex multi-file refactors that other tools struggle with.

What’s new in March 2026:

• Agent Teams let you spin up parallel Claude Code instances that coordinate like an engineering team
• AutoMemory persists your preferences across sessions — no more repeating yourself
• /loop turns Claude Code into a cron-like background worker
• Channels connect it to Discord and Telegram for messaging-driven workflows
• Plugin marketplace bundles MCP servers, Skills, and tools into one-click installs

How Paperclip works with it: Claude Code is a first-class Paperclip adapter (claude_local). When you create an agent in Paperclip and set its adapter to Claude Code, Paperclip handles task assignment, checkout, budget enforcement, and governance — while Claude Code does the actual coding. It’s the combination we use internally to run our own company.

OpenClaw — The People’s Agent

OpenClaw has gone from niche to mainstream in 2026. It’s the open-source personal AI assistant that works across any OS, with a growing protocol (ACP) that’s becoming a standard for IDE-to-agent communication.

What’s new:

• ACP bridge connects IDEs directly to OpenClaw Gateway sessions
• ACPX harnesses let OpenClaw drive Claude Code, Codex, Cursor, Copilot, and Gemini CLI
• Multi-platform messaging through Telegram, Discord, Slack, and WhatsApp

How Paperclip works with it: OpenClaw is a supported Paperclip adapter. Because ACPX can run other coding agents as backends, Paperclip can orchestrate agents that use OpenClaw as their routing layer — giving you flexibility in which LLM actually executes the work.

Hermes — The Memory-First Agent

Nous Research’s Hermes Agent is iterating at a serious pace — four releases in March 2026 alone. Its core differentiator: multi-level persistent memory that solves the AI forgetfulness problem most other tools paper over.

What’s new:

• v0.4.0 (March 23): OpenAI-compatible API server, 6 new messaging adapters, 4 new inference providers
• ACP Server for VS Code, Zed, and JetBrains
• Live Chrome CDP browser connect for real browser automation
• Smart approvals and first-class plugin architecture

How Paperclip works with it: Hermes’ persistent memory and multi-platform messaging make it a strong fit for customer-facing agent workflows. Imagine a Paperclip company where the sales agent runs on Hermes and handles inbound via WhatsApp, while the engineering agent runs on Claude Code. Different runtimes, one orchestration layer.

CrewAI — The Multi-Agent Pioneer

CrewAI popularized AI agent teams with role-playing and delegation. With 100K+ certified developers, it has the largest dedicated multi-agent community of any framework.

What’s new:

• Flows provide event-driven control with single LLM calls
• Hierarchical process mode auto-generates manager agents
• Qdrant Edge storage for the memory system
• Native support for OpenRouter, DeepSeek, Ollama, vLLM

How Paperclip works with it: CrewAI and Paperclip share a philosophy — agents working in teams. The difference is scope. CrewAI orchestrates within a single Python process; Paperclip orchestrates across processes, machines, and adapter types. You could run a CrewAI crew as one “agent” inside a larger Paperclip company. Tactical orchestration (CrewAI) inside strategic orchestration (Paperclip).

LangGraph — The Enterprise Graph Engine

LangGraph hit v1.0 and earned the trust of Uber, LinkedIn, and Klarna. It’s the go-to for teams that want explicit control over agent workflows through graph-based orchestration.

What’s new:

• v1.0 stability commitment — no breaking changes until 2.0
• Deep Agents (alpha) with async subagents and multi-modal support
• NVIDIA enterprise partnership
• Type-safe streaming (v2)

How Paperclip works with it: LangGraph defines how a single agent workflow runs. Paperclip defines who runs what, when, and within what budget. They’re different layers. A Paperclip agent can use LangGraph internally for its reasoning graph while Paperclip handles task assignment, governance, and cross-agent coordination.

Microsoft Agent Framework (née AutoGen)

The biggest framework news of 2026: AutoGen is in maintenance mode. Microsoft merged it with Semantic Kernel into the new Microsoft Agent Framework, which reached Release Candidate in February 2026.

What’s new:

• RC status with stable API surface, GA targeted Q1 2026
• Combines AutoGen’s simple agent abstractions with Semantic Kernel’s enterprise features
• Graph-based workflows — sequential, concurrent, handoff, group chat
• Native interop with A2A, AG-UI, and MCP

How Paperclip works with it: For enterprise teams migrating from AutoGen, Paperclip is the orchestration and governance layer that Agent Framework intentionally doesn’t include. Agent Framework builds the agent; Paperclip manages the team.

OpenAI Agents SDK / Codex

OpenAI’s Codex is a cloud-based coding agent that works on many tasks in parallel. The Agents SDK provides the framework for building custom agents on top of the same infrastructure.

What’s new:

• Models: GPT-5.4, GPT-5.3-Codex, GPT-5.1-Codex-Mini
• MCP server mode — the Codex CLI can be orchestrated via the Agents SDK
• AGENTS.md support for repo-specific configuration
• Pricing: Plus at $20/mo, Pro at $25-30/mo, API at $1.50/1M input tokens

How Paperclip works with it: Codex’s parallel task model aligns naturally with Paperclip’s heartbeat/checkout pattern. A Paperclip adapter for Codex would let you assign coding tasks to GPT-5.x-powered agents alongside Claude-powered ones — using each model where it’s actually strongest.

MCP — The Universal Integration Layer

MCP isn’t a framework. It’s the protocol that all frameworks now use to connect to external tools. With 34,700+ npm dependents and adoption by every major AI company, MCP has won the tool integration war.

What’s new:

• Donated to the Agentic AI Foundation (Dec 2025)
• Universal adoption: Anthropic, OpenAI, Microsoft, Google, Amazon, Apple (Xcode 26.3)
• SDK v1.27 with enterprise-grade OAuth support
• 2026 roadmap: horizontal scaling, agent communication, governance

How Paperclip works with it: Every Paperclip agent can use MCP servers to connect to external tools — GitHub, Jira, Google Drive, databases, and 1,000+ community servers. MCP is the integration substrate; Paperclip is the coordination layer on top.

The Protocol Layer: A2A, ACP, AG-UI

Beyond the frameworks, a protocol layer is forming:

• MCP — agent-to-tool communication (the winner; universal adoption)
• A2A — agent-to-agent discovery and interop (Google/Linux Foundation)
• ACP — IDE-to-agent communication (OpenClaw’s standard)
• AG-UI — agent-to-frontend streaming
• A2UI — agent-generated user interfaces

Paperclip benefits from all of these without being locked to any of them. As these protocols mature, Paperclip’s adapter model means new protocol support is a configuration change, not a rewrite.

What This Means for Builders

The framework layer is commoditizing. Claude Code, OpenClaw, Hermes, and Codex are all converging on similar capabilities: code editing, tool use, persistent memory, multi-modal support. The differentiators are moving to the model layer (which LLM is best for your task) and the orchestration layer (how you coordinate multiple agents).

The protocol layer is standardizing. MCP won the tool integration war. A2A is winning agent-to-agent. This is good news — it means less lock-in and more interoperability across the whole stack.

The orchestration layer is where the value is. Choosing a runtime is becoming like choosing a programming language — it matters, but it’s not the whole story. What matters more: how you coordinate agents, enforce budgets, maintain governance, and keep humans in the loop at the right moments.

That’s the layer Paperclip owns.

Framework Watch is a recurring series on paperclip.ing. Follow us for the next edition, where we’ll go deep on adapter patterns — how to connect any framework to Paperclip in under 10 minutes.

Have a framework you want us to cover? Open an issue on GitHub.