Harnesses

A harness is the runtime that wraps your agent — the orchestration layer that decides when to call a tool, what to send the model, how to assemble a response. CodeNow ships first-class scaffolds for four harnesses out of the box.

Shipped today

Anthropic’s official harness. Best Claude integration. Streaming query() API.

Claude Agent SDK (TS + Python)

Cursor’s harness with Composer 2 + multi-model. Sandboxed cloud VMs, subagents, hooks.

Cursor SDK (TypeScript)

Microsoft’s official agent SDK (Semantic Kernel + AutoGen merged). Azure OpenAI native.

Microsoft Agent Framework

Google’s Agent Development Kit. Optimized for Gemini 2.0, Vertex AI native.

Google ADK

Coming

LangGraph (LangChain ecosystem), OpenAI Agents SDK, CrewAI, MCP server. Track progress in the agent-kits.js source .

Picking a harness

If your stack is…	Pick
Anthropic API, want best Claude integration	Claude Agent SDK
Already using Cursor / want multi-model + cloud sandboxing	Cursor SDK
Microsoft enterprise (Windows, Office, Azure, .NET)	Microsoft Agent Framework
Google Cloud / Vertex AI / Gemini-native	Google ADK
Don’t care, just give me the fastest first-run	Claude Agent SDK (TypeScript) — the default Hello Agent kit

CodeNow’s Live Agent Replay and manifest work the same regardless of harness. Switch by re-scaffolding with a different kit; the lifecycle wrapper around it stays identical.

How CodeNow integrates each harness

Each scaffold drops a recorder.ts (or recorder.py) into your project that wraps the harness’s stream. The wrapper writes JSONL events to .codenow/runs/{slug}/ so the Runs tab can replay any past invocation. Zero runtime dependencies — Node built-ins / Python stdlib only. Fails open: any IO error is swallowed so a recorder bug never breaks your agent.

This means every harness gets the same observability surface — Replay, Cost tracking, Fork-to-current-code, audit trail — for free.