Engineered for Reasoning,Built for Production.
A deterministic runtime for LLM agents — typed tools, provider fallback, and guardrails that can't be skipped. Built in raw TypeScript, no framework lock-in.
Every capability, verified.
Guardrails & Approval Gate
Halt execution for human review when high-risk tools are called or budgets are exceeded.
Multi-Provider Fallback
Seamlessly failover from Claude to OpenAI if rate limits or outages occur, without changing code.
Structured Output & Self-Repair
Enforce strict schema compliance. The runtime automatically prompts the model to fix malformed JSON.
Multi-Agent Handoffs
Route tasks between specialized agents while maintaining a unified conversational state and memory.
Memory & Sessions
Built-in episodic and factual memory systems to persist context across long-running interactions.
Tracing & Observability
Export OpenTelemetry traces for every tool call, generation step, and state transition.
Five lines to a production agent.
Stop writing bespoke while-loops and fragile parsing logic. The SDK provides a fully managed state machine that handles the complexities of LLM orchestration — from multi-step tool use to schema repair and human-in-the-loop approvals.
import { Agent, AnthropicProvider, Memory } from '@weave-agent/core';
const agent = new Agent({
provider: new AnthropicProvider({ model: 'claude-3-5-sonnet' }),
memory: new Memory.SessionMemory(),
tools: [searchTool, dbTool],
guardrails: {
approval: ['dbTool.dropTable']
}
});
// The runtime handles tool execution and validation automatically
const result = await agent.run("Find the latest user and drop their table");