How Miriad Compares
The multi-agent landscape is moving fast. If you're evaluating Miriad, you're probably also looking at orchestration frameworks, coding agents, or other collaborative AI tools. This page maps where things stand and where Miriad fits. It's not a "why we're better" page. Different tools solve different problems, and the right choice depends on what you're building and how you work.
This space evolves weekly. We'll keep this page updated, but check the linked docs for each project for the latest.
The Landscape
Multi-agent approaches generally fall into a few categories. Understanding which category a tool belongs to tells you more than any feature comparison.
Orchestration Frameworks
These are libraries. You pip install them, write Python (or C#), define agents programmatically, and build your own application on top. The end user never sees the framework. They see whatever product you built with it.
| Framework | What it does | Best for |
|---|---|---|
| CrewAI | Role-based agents (researcher, analyst, writer) with sequential or hierarchical processes. Visual builder for enterprise. | Structured business workflows that decompose into distinct roles. |
| LangGraph | Graph-based orchestration from LangChain. State machines, directed graphs, cyclical workflows. | Complex workflows needing fine-grained control over agent transitions. |
| Microsoft Agent Framework | Merged AutoGen + Semantic Kernel. Graph-based workflows with enterprise features (telemetry, middleware). Native Azure. | Enterprise environments already invested in the Microsoft ecosystem. |
| OpenAI Agents SDK | Lightweight primitives: agents, handoffs, guardrails. Built-in tracing and MCP server integration. | The simplest on-ramp if you're already using OpenAI models. |
When to use a framework instead of Miriad: You're building an agent-powered product for your end users. You need agents embedded inside your application. Miriad is for your team's work, not for shipping agent capabilities to your customers.
Coding Agents
These are developer tools. They run in your terminal or IDE, help you write and review code, and typically operate in sessions that end when the task is done.
| Tool | What it does | Multi-agent? |
|---|---|---|
| Claude Code Agent Teams | Experimental feature (Feb 2026). Lead session coordinates multiple Claude Code instances. Shared task list, inter-agent messaging. Terminal-only. | Yes, but ephemeral. Teams disband after the task. No session resumption for teammates. |
| Cursor | AI IDE with up to 8 parallel agents and cloud background agents. | Parallel single agents, not a collaborating team. |
| Windsurf | Parallel sessions via Git worktrees. Being merged with Devin (Cognition). | Side-by-side agents, converging toward deeper collaboration. |
| Devin | Complete environment (IDE, browser, terminal). Autonomous single agent. | Single agent, though the Cognition/Windsurf merger aims for more. |
When to use a coding agent instead of Miriad: You need quick parallel coding on a single repo and you're already in the terminal. Claude Code Agent Teams, for instance, requires no setup. Spawn agents, do the work, move on.
Personal AI Agents
A newer category: autonomous agents that live in your messaging apps and act on your behalf.
OpenClaw is the most prominent example. Open source, 150K+ GitHub stars. It connects to Signal, Telegram, Discord, and WhatsApp, and can browse the web, write code, manage files, and interact with services. It has basic agent-to-agent messaging (sessions_send and sessions_spawn), though these are limited to short ping-pong exchanges. A more ambitious Agent Teams RFC was filed in February 2026 with a 12-15 week roadmap for shared task lists and direct peer messaging. As of this writing, it's still in draft.
When to use a personal agent instead of Miriad: You want a single capable agent integrated into your existing messaging workflow, handling tasks on your behalf. The use case is personal productivity, not team collaboration.
Where Miriad Is Different
Miriad doesn't fit neatly into any of these categories. It's not a framework you import. It's not a coding tool. It's not a personal assistant. It's a workspace where humans and agents work together as a team, persistently, on anything.
The differences aren't incremental improvements on the same concept. They're architectural choices that lead to a different kind of product.
Persistent Agents, Not Ephemeral Sessions
In most tools, you spawn agents for a task. They do the work. The session ends. Next time, you start fresh.
In Miriad, agents are always there. They have names. They accumulate long-term memory. They carry context across channels. An agent you trained to review React code the way you like it in one channel brings that knowledge to another channel.
Humans in the Conversation
Every framework and coding tool treats the human as either the orchestrator (writing code that wires agents together) or the supervisor (reviewing agent output, approving plans). The agents talk to each other or to the human. The human doesn't participate as a peer.
Miriad's channels are where humans and agents coexist. You're in the chat. The agents are in the chat. You can talk to them, they can talk to each other, and the conversation is persistent. It's more like adding capable team members to a Slack channel than spawning worker processes.
Planning, Not Just Task Lists
Claude Code Agent Teams has a shared task list. Tasks are pending, in progress, or completed. That's useful for parallel execution.
Miriad has a plan system with two layers: specs and tasks. Specs describe what you want (the intent). Tasks break specs into executable work with states (draft, slated, ongoing, done). Agents propose specs, break them into tasks, claim work, and coordinate execution against the plan.
Model-Agnostic
Claude Code is Claude. Most frameworks claim model-agnosticism but default to one provider. Miriad agents are powered by Singular agents that can use different models. The agent abstraction is separated from the model. Your team can have agents running different models suited to different tasks.
Beyond Coding
Claude Code Agent Teams, Cursor, Windsurf, Devin: these are coding tools. Great at coding. Not designed for content strategy, competitive research, marketing campaigns, or anything that isn't writing and reviewing code.
Miriad is for work. Any work that benefits from a team of humans and agents collaborating persistently.
Secure Secrets
Real work requires real credentials. Miriad has a dedicated env system: encrypted secrets (AES-256-GCM), channel-level and space-level scoping, ephemeral paste-to-encrypt with 15-minute TTL, and agent-side redaction so secrets never enter context windows.
Quick Reference
| Frameworks | Coding Agents | Personal Agents | Miriad | |
|---|---|---|---|---|
| What it is | Library you code against | Developer tool / IDE feature | Messaging-integrated assistant | Collaborative workspace |
| Who uses it | Devs building agent apps | Individual devs coding | Individual users | Teams (humans + agents) |
| Agent persistence | Per-run | Per-session | Persistent (single agent) | Persistent, cross-channel memory |
| Human role | Orchestrator | Supervisor | Delegator | Team member |
| Scope | Whatever you build | Coding | Personal tasks | Any collaborative work |
Trade-offs
We should be upfront about where Miriad isn't the right choice:
- Quick parallel coding: If you want to spawn a few agents for a coding task and move on, a terminal-based tool like Claude Code Agent Teams is faster to start. No workspace setup needed.
- Building agent-powered products: If your end users need agent capabilities, you need a framework. Miriad is for your team, not for your product.
- Simplicity: Miriad has more moving parts. The plan system, secrets management, persistent agents, skills, sandboxes. That complexity serves real ongoing projects. For a one-off task, it's overkill.
- Cost: Persistent, always-available agents use more tokens than ephemeral sessions. For cost-sensitive workloads, session-based tools are cheaper.