Research — March 2026

Edge-native agentic frameworks

A deep comparison of six TypeScript agentic frameworks, evaluated for edge cloud deployment — one agent instance per user, running at the network edge with durable state.

swarm.ing research | 6 frameworks compared | Updated March 30, 2026
Overview
Vercel AI SDK
Vercel · v6 · 20M+ monthly downloads
TS
Core abstractionAgent class + tool loop
ExecutionStateless per-request
StateExternal (BYO storage)
Multi-agentManual via tool handoff
Providers40+ unified API
MCPClient (connect to servers)
UI streamingSSE + React/Vue/Svelte hooks
LicenseApache 2.0
Mastra
Ex-Gatsby team · v1 · $13M seed · 150K wk DLs
TS
Core abstractionAgent + Workflow graph
ExecutionGraph-based DAG workflows
StateBuilt-in storage + memory
Multi-agentSupervisor pattern, networks
Providers40+ (via Vercel AI SDK)
MCPClient + server
UI streamingSSE + CopilotKit/Cedar-OS
LicenseApache 2.0 / BUSL (EE)
Inngest AgentKit
Inngest · Durable orchestration engine
TS
Core abstractionAgent networks + routers
ExecutionDurable step functions
StateAutomatic (step-level)
Multi-agentNative networks + KV state
ProvidersOpenAI, Anthropic, Gemini+
MCPNative MCP as tools
UI streaminguseAgent React hooks
LicenseApache 2.0 / SaaS
Claude Agent SDK
Anthropic · Renamed from Claude Code SDK · 3.5M+ DLs
TSPY
Core abstractionClaude Code harness
ExecutionSubprocess (Claude CLI)
StateSession transcripts on disk
Multi-agentSubagents (fg/bg)
ProvidersClaude only
MCPFull (client + custom tools)
UI streamingEvent iterator / SSE
LicenseCommercial ToS
PersistentAI / FireFlow
PersistentAI · DBOS · Tetra/TON L2
ChainVisual
Core abstractionVisual flow nodes
ExecutionDBOS durable workflows
StatePostgreSQL + LakeFS VFS
Multi-agentSwarm delegation
ProvidersOpenAI, Claude, Gemini, local
MCPNative (auto node gen)
UI streamingtRPC + WebSocket (Effector)
LicenseBUSL-1.1
Edge readiness — one agent per user at the edge

Scoring each framework on the key requirement: can I spin up a separate, stateful agent instance per user on the network edge with zero infrastructure management?

9.5
CF Agents SDK
Native DO per-user, SQLite, hibernation
6.5
Mastra
CF deployer exists, but not DO-native
5.0
Vercel AI SDK
Stateless — edge compute, not edge state
4.5
Inngest AgentKit
Durable but centralized, not edge-native
2.0
Claude Agent SDK
Requires CLI subprocess — server only
2.0
PersistentAI
Docker/self-hosted — no edge story
Edge capabilityVercel AI SDKMastraInngestClaude AgentCF AgentsPersistentAI
Instance-per-user Stateless ~ Via external storage ~ Via step state ~ Per-session files DO per user/session ~ Per-workspace
Co-located compute + state ~ Disk-local Same-thread SQLite ~ PostgreSQL (not co-located)
Hibernation (zero cost when idle) N/A Hibernatable WebSockets
Global edge locations Vercel Edge Via CF deployer Centralized Server only 300+ cities Docker host
Scale to millions of instances N/A ~ Depends on backend ~ Via cloud tier Millions of DOs
WebSocket per agent SSE only SSE only ~ Via useAgent Built-in per-DO tRPC WebSocket
Scheduling / cron per agent ~ Via workflow Inngest cron DO alarms + schedule() DBOS scheduled
Architecture comparison
DimensionVercel AI SDKMastraInngestClaude AgentCF AgentsPersistentAI
Durability Statelessuse workflow add-on ~ Suspend/resumehuman-in-the-loop Step checkpointssurvives crashes ~ Session filesfile checkpointing DO + Workflowsauto-retry, hibernate DBOS exactly-oncetime-travel debug
Storage model BYO (KV, Postgres…) Built-in memory (short + long) Inngest cloud state + KV Filesystem transcripts SQLite per-DO + KV/D1/R2 PostgreSQL + LakeFS VFS
Multi-agent Tool handoffs (manual) Supervisor + parallel/branch Networks + typed state router Subagents (fg/bg) Cross-DO RPC + Workflows Swarm delegation + messaging
Observability Vercel Observability Built-in evals + tracing Dev server traces + dashboard Event stream + usage metrics Workers Logs + AI Gateway Flow debugger + breakpoints
Blockchain / payments Tetra L2 + x402 + NFT agents
Human-in-the-loop Tool approval (v6) Workflow suspend Durable steps AskUserQuestion tool waitForEvent DBOS signal await
Developer experience & deployment
DimensionVercel AI SDKMastraInngestClaude AgentCF AgentsPersistentAI
Getting started npm i ai npx create mastra@latest npm i @inngest/agent-kit npm i @anthropic-ai/claude-agent-sdk npm create cloudflare -- --template agents-starter git clone + pnpm + docker
Local dev Next.js dev server Mastra Studio (:4111) Inngest Dev Server Node.js / CLI direct wrangler dev Docker Compose
Deploy target Vercel, any Node host Vercel, CF, standalone Any host + Inngest Cloud Any server with CLI Cloudflare Workers Docker / self-hosted
Vendor lock-in Low Low Medium High (Claude only) Medium (CF infra) Low (self-hosted)
Maturity / traction Very high (20M/mo) High (150K/wk, enterprise) Medium (growing) Medium (rapid v0.2.x) Medium (GA) Early (pre-traction)
Cost model Free SDK + hosting Free OSS + Cloud Free tier + usage Free SDK + API tokens Free tier + DO/Workers Infra costs only
Verdict — for edge-first agent infrastructure
Best edge fit
Cloudflare Agents SDK
The only framework that natively delivers one stateful agent instance per user at the edge. Each DO has co-located SQLite, WebSockets, scheduling, and hibernation. Scales to millions. Pairs with your existing CF stack (D1, Vectorize, R2, Dynamic Workers).
Best orchestration layer
Mastra
Richest workflow primitives — supervisor multi-agent, graph-based DAGs, built-in evals and memory. Ships @mastra/deployer-cloudflare for CF Workers deployment. Use as the orchestration brain that runs inside your CF Agent DOs.
Highest DX leverage
Claude Agent SDK
Unmatched for coding agents — Claude Code's full harness (Bash, Read, Edit, Glob) exposed as SDK. Model-locked to Claude but nothing else comes close for dev tooling agents. Use for internal dev workflows, not user-facing edge agents.
Worth watching
PersistentAI
Interesting for Telegram Mini App + blockchain settlement. DBOS exactly-once execution is real. x402 micropayments + agent NFTs is novel. But early-stage, BUSL license, self-hosted only. No edge deployment story yet.