Why Halo?

Comparing Halo with other approaches to building AI agents.

Halo vs Raw API Calls

// ❌ Raw API: manual cache management, DIY loop, no tracking
const messages = [
  { role: "system", content: "You are..." },
  ...toolsAsMessages,
  ...history,
];
const res = await fetch("https://api.deepseek.com/chat/completions", {
  body: JSON.stringify({ messages, tools }),
});
// Is caching working? No idea.
// Tool calls returned? Need to write the loop yourself.
// Cost tracking? Manually from usage fields.

// ✅ Halo: automatic cache, agent loop, tracked stats
const agent = halo.agent({ system: "You are...", tools: { ... } });
const result = await agent.run("Do the task");
console.log(agent.stats.caching?.cacheHitRate); // 0.92
console.log(agent.stats.caching?.estimatedSavingsUsd); // $0.023

Halo vs Vercel AI SDK

	Halo	Vercel AI SDK
Focus	Cache-first agent runtime	Provider-agnostic LLM toolkit
Caching	Built-in StablePrefix, auto-managed	Manual via experimental_cache options
Agent Loop	agent.run() — auto tool execution	Manual with maxSteps + tool handling
Cost Tracking	agent.stats.caching built in	Manual from usage
Provider Model	Adapters with cache strategies	Provider packages
SSE Protocol	Compatible via toDataStream()	Native 0:/d: protocol
Best For	Agent-heavy apps, cost-sensitive, multi-turn	General LLM apps, broad provider support

The two are complementary — Halo's toDataStream() makes it compatible with Vercel AI SDK's useChat().

Halo vs LangChain

	Halo	LangChain
Philosophy	Minimal, cache-first	Maximal, abstraction-heavy
Bundle Size	~5KB per package	500KB+
Caching	Automatic prefix cache	Manual LCEL caching
Learning Curve	~5 minutes	Hours to days
Debugging	Direct adapter code, visible HTTP	Deep chain abstraction

When to Choose Halo

• You're building multi-turn agents with tool calling
• You care about token costs and want automatic savings tracking
• You use DeepSeek (or plan to support multiple providers)
• You want a minimal, auditable codebase (zero external dependencies in core)
• You're shipping to production and need keep-alive, stats, and error events

When Not to Choose Halo

• You need structured output generation (JSON mode, objects) — use Vercel AI SDK
• You need embeddings, image generation, or speech — use provider SDKs directly
• You're doing a single-shot prompt with no conversation — raw API is fine