agent-subsystems

Route each step of a multi-agent task to the right AI model automatically.

The problem

When you build a pipeline where research goes to a cheap fast model, code review goes to a powerful one, and planning goes to the most capable — you end up writing the same routing logic, context translation, and state handoff boilerplate every time. There's no standard way to describe a task's cost/capability tradeoff and have a runtime enforce it. Each new workflow is a fresh engineering problem.

Quick start

npm install -g interchange

export ANTHROPIC_API_KEY=sk-ant-...
export OPENAI_API_KEY=sk-...

interchange run "Research recent advances in RAG, then write a concise summary"

Install

# CLI
npm install -g interchange

# Library (types, adapters, state management)
npm install @npow/interchange-core

Usage

Run a task — automatically routed to the optimal model

interchange run "Refactor the auth module to use JWT"

Force a specific role

interchange run-as coder "Add rate limiting to the API"
interchange run-as reviewer "Review the changes in auth.ts"

Resume from a previous handle

# Save the handle from a previous run
HANDLE=$(interchange run "Research X" --json | jq -r .handle)
interchange resume "$HANDLE"

Use @interchange/core in your own code

import { translateRecord, StateManager, MemoryBackend } from "@npow/interchange-core";

// Translate Claude tool calls to Codex JSONL format
const result = translateRecord(claudeRecord);

// Manage task state across agent handoffs
const state = new StateManager(new MemoryBackend());
const task = await state.create({ taskId: "t1", role: "coder", goal: "Fix the bug" });

How it works

interchange run decomposes your task into a DAG of subtasks. Each node is classified by required capability (planning, coding, research, review) and routed to the best available model within optional cost constraints. The baton protocol carries structured context — facts, decisions, tool call history — across handoffs as a typed system prompt. State is persisted to ~/.interchange/state.json.

Built on the Vercel AI SDK with support for Anthropic, OpenAI, and Google models.

Configuration

Pass --max-cost <usd-per-1k> to run / run-as to cap model selection by cost.

Development

git clone https://github.com/npow/interchange
cd agent-subsystems
npm install
npm run build --workspaces
npm test --workspaces

Packages:

• packages/core/ — @interchange/core: canonical types, format adapters, state management
• packages/interchange/ — interchange: router, orchestrator, CLI

License

Apache 2.0