Always
LLM tokens
Every framework calls a language-model API and you pay that vendor for the tokens consumed. Pay-as-you-go, identical whether you choose Anthropic, OpenAI, Bedrock, Azure OpenAI, Gemini, or a local model via Ollama.
Agent Framework Harness — Wiseway Quoting & Transport
What builds the event-driven assistants.
The Knowledge Assistant uses a chat app. Quoting and Transport need a different layer — a framework that lets your code receive an incoming email, run an agent, call tools, and write the result back into your TMS. Here are the credible contenders as of May 2026.
How pricing works
Every framework is open source. The real cost is who hosts it.
Software licence isn't the variable. Where you run it is.
Three cost models: pay-per-token (always), self-hosted infrastructure, vendor-hosted per-session.
How pricing works
Every framework is open source. The real cost is who hosts it.
Software licence isn't the variable. Where you run it is. Three cost models: pay-per-token (always), self-hosted infrastructure, vendor-hosted per-session.
Self-hosted
Your infrastructure
Open-source frameworks run in your process or container. Cost is your Azure / AWS bill — for Wiseway, that's roughly AU$50–200/month for a low-volume agent service in AU East. No per-seat fees, no per-session fees.
Vendor-hosted
Per-session pricing
Anthropic's Managed Agents (currently in beta) charges US$0.08 per active session-hour on top of token costs. They run the sandbox, you pay only while the agent is actively executing.
Side-by-side
Ten frameworks, the dimensions that decide.
Eight production frameworks plus two chat-surface SDKs flagged for completeness.
HITL approvals, durable execution, observability, model-agnosticism — side by side, sourced and dated.
Side-by-side
Ten frameworks, the dimensions that decide.
Eight production frameworks plus two chat-surface SDKs flagged for completeness. HITL approvals, durable execution, observability, model-agnosticism — side by side, sourced and dated.
| Framework | Languages | Model agnostic |
MCP | HITL approval |
Durable execution |
Observ- ability |
Maturity | Hosting |
|---|---|---|---|---|---|---|---|---|
| Production agent frameworks | ||||||||
| Claude Agent SDK Anthropic · MIT · Python + TS | Python, TypeScript | ✗Claude only | ✓ | ⚠via hooks | ⚠session forking | ⚠custom hooks | Stable | Self-hostruns in your process |
| Claude Managed Agents Anthropic · hosted REST API | Any (REST) | ✗Claude only | ✓ | ⚠steering, beta | ✓stateful sessions | ✓built-in | Public beta | Vendor-hostedUS $0.08 / session-hour + tokens |
| LangGraph LangChain Inc. · MIT | Python, TypeScript | ✓ | ✓ | ✓first-class | ✓checkpointer | ✓LangSmith | Stable | Self-hostor LangGraph Platform (paid) |
| Mastra Mastra Inc. · MIT · TS-first | TypeScript | ✓ | ✓ | ✓ | ✓workflow steps | ✓OpenTelemetry | 1.0 (Jan 2026) | Self-hostor Mastra Cloud (paid) |
| OpenAI Agents SDK OpenAI · MIT (replaces Swarm) | Python, TypeScript | ⚠via LiteLLM only | ✓ | ⚠guardrails | ⚠sessions only | ✓OpenTelemetry | Stable | Self-hosttight coupling to OpenAI Responses API |
| Microsoft Agent Framework Microsoft · MIT · AutoGen successor | Python, .NET | ✓ | ✓ | ✓ | ✓checkpointing | ✓built-in | 1.0 GA | Self-hostor Azure AI Foundry |
| Pydantic AI Pydantic team · MIT | Python | ✓ | ✓ | ✓ | ✓durable agents | ✓Logfire | Stable | Self-hosttype-safe, no vendor option |
| CrewAI CrewAI Inc. · MIT | Python | ✓ | ✓ | ⚠callbacks only | ⚠limited | ⚠AMP for enterprise | Stable | Self-hostor CrewAI AMP (paid control plane) |
| Chat-surface SDKs — listed for completeness, not full agent frameworks | ||||||||
Goose goosed
AAIF · Apache 2.0 · REST + SSE
|
Any (REST) | ✓ | ✓ | ✗ | ⚠session-only | ⚠basic | Embeddable | Self-hostruns the goosed daemon |
| LibreChat Independent · MIT · chat surface | n/a | ✓ | ✓ | ✗ | ✗ | ⚠basic | Chat client only | Self-hostuse alongside LangGraph/Mastra |
Tick legend: ✓ capability present, ⚠ partial / caveat, ✗ not supported.
The "no vendor lock-in" takeaway: Claude Agent SDK and Claude Managed Agents are Claude-only on the model side. If model neutrality is non-negotiable, the most credible alternatives are LangGraph (Python or TS, incumbent, used by Klarna, Uber, LinkedIn), Mastra (TypeScript, 1.0 production-ready since Jan 2026), or Microsoft Agent Framework (Python or .NET, GA, but pulls toward Azure). The Goose and LibreChat rows are there to be honest — neither is a true framework competitor; you'd pair either with LangGraph or Mastra as the brain.
Sources: Claude Agent SDK docs · Claude Managed Agents docs · LangGraph · Mastra · OpenAI Agents SDK · Microsoft Agent Framework · Pydantic AI · CrewAI · Goose · LibreChat Agents.
Mapped to your brief
What Wiseway specifically asked for.
Seven requirements from your brief — and which contenders satisfy each one.
Quoted lines from sections 4–6 of the Wiseway AI Brief, mapped to framework recommendations.
Mapped to your brief
What Wiseway specifically asked for.
Seven requirements from your brief — and which contenders satisfy each one. Quoted lines from sections 4–6 of the Wiseway AI Brief, mapped to framework recommendations.
| Your requirement | What it demands | Best fits |
|---|---|---|
| "Important — never auto-send. A sales rep always reviews and sends. Wrong rates to clients like SHEIN cost us real money." | First-class human-in-the-loop / approval primitives — the agent drafts, a human sends. | LangGraph · Mastra · Pydantic AI · Microsoft Agent FrameworkClaude Agent SDK can do it via hooks but it's not first-class. |
| "Flag quotes below margin thresholds for manager approval." | Conditional escalation steps inside the workflow — branch to manager when margin < threshold. | LangGraph (graph approval nodes) · Mastra (workflow steps) · Microsoft Agent Framework |
| "Log everything to our CRM." + "Audit trail: every write carries the source message ID." | Built-in observability + structured event logging — every tool call traceable to its triggering message. | LangGraph + LangSmith (gold standard) · Mastra (OpenTelemetry) · Microsoft Agent Framework · Pydantic AI + Logfire |
| "Multi-jurisdiction — gives the right answer for AU vs NZ vs USA vs UK staff." | Conditional branching logic and per-jurisdiction skill routing. | LangGraph (graph-native branching) · Microsoft Agent Framework · CrewAI |
| "Don't lock us into a single vendor or model provider." | Truly model-agnostic — works with Anthropic, OpenAI, Bedrock, Azure OpenAI, Gemini, local. | LangGraph · Mastra · Pydantic AI · Microsoft Agent Framework · CrewAIRules out Claude Agent SDK and Claude Managed Agents on this axis alone — both are Claude-only on the model side. |
| "We want to own the outcome — maintain and extend what's been built." | Open source, Wiseway-owned skills/workflows, no proprietary runtime. | Every framework above except Managed Agents — all open source under MIT / Apache. Managed Agents is hosted, so the runtime stays with Anthropic. |
| "Pragmatic delivery — pilot small, prove value, expand." | Strong developer experience, fast time-to-first-working-agent, low operational floor. | Claude Agent SDK (best DX if Claude-only is OK) · Mastra · Pydantic AI |
Pattern that satisfies all seven rows simultaneously: LangGraph (Python) or Mastra (TypeScript) as the agent runtime, with MCP servers for SharePoint / TMS / CargoWise integration, deployed in Azure AU East. Both give you HITL, durable execution, observability, and model-agnosticism out of the box — exactly the items the brief flags as non-negotiable. Claude Agent SDK remains the easiest on-ramp if you accept the Claude-only constraint and add HITL via hooks.