feat: migrate to createAgent API and middleware architecture

[christian-bromann]

Overview

This PR migrates the Deep Agents library from a custom LangGraph-based setup to the new langchain createAgent API with a middleware-driven architecture. It simplifies state management, standardizes agent creation, and aligns the project with current LangChain best practices.

Why

• Reduce bespoke state/schema plumbing in favor of built-in agent capabilities.
• Adopt a cleaner, composable middleware model for filesystem and todo operations.
• Align with the latest langchain alpha APIs to enable future enhancements.

What Changed

• Architecture

  • Replaced createReactAgent with createAgent.
  • Introduced middleware-based design (fsMiddleware, todoMiddleware).
  • Removed custom state scaffolding in favor of runtime middleware state.

• Code

  • Deleted: src/state.ts, src/model.ts, src/interrupt.ts.
  • Moved tools to middleware: src/tools.ts → src/middleware/fs.ts; added src/middleware/todo.ts; new src/middleware/index.ts.
  • Refactored src/subAgent.ts:
    • Uses createAgent and middleware.
    • model can be a string or LanguageModelLike, defaulting to "openai:gpt-4o-mini".
    • Removed internal hardcoded/built-in tools map; resolution now relies solely on provided tools.
    • When subagent.tools is unspecified, all provided tools are used.
  • Updated src/types.ts:
    • Added SubAgentSchema (Zod) and SubAgent type based on it.
    • Simplified CreateDeepAgentParams; interruptConfig now uses HumanInTheLoopMiddlewareConfig["toolConfigs"].
    • Simplified CreateTaskToolParams generic to z.ZodObject.
  • Config and examples:
    • tsconfig.json: fix extends path to @tsconfig/recommended/tsconfig.json.
    • examples/research/research-agent.ts: updated to new API.

• Dependencies

  • Bumped ESLint/TypeScript-related deps.
  • Upgraded to langchain alpha and latest related packages (@langchain/langgraph@next, @langchain/anthropic@alpha).
  • Updated lockfile accordingly.

Breaking Changes

• Removal of DeepAgentState and custom reducers (src/state.ts).
• Removal of built-in tools registry inside subAgent.ts:
  • Callers must pass tools explicitly (by name via subagent.tools mapping into the provided tools object).
• createTaskTool signature changes:
  • stateSchema no longer required.
  • model accepts either a model string or a LanguageModelLike.
• Interrupts/state handling now flows through middleware; any prior usage of interrupt.ts patterns must be updated.

Migration Guide

• Replace any createReactAgent usage with createAgent.
• Pass tools explicitly via tools: Record<string, StructuredTool>; if a subagent specifies no tools, all provided tools will be used.
• Remove dependencies on DeepAgentState; move stateful behaviors into middleware or rely on the runtime state the agent manages.
• Update types/imports to use SubAgentSchema / SubAgent and the simplified CreateDeepAgentParams.
• Ensure your model is provided as a string (e.g., "openai:gpt-4o-mini") or as a LanguageModelLike.

[christian-bromann]

Unfortunately this is blocked again until we do another alpha release.

[amitchaudharyc]

Hey @christian-bromann , glad to see the changes. But can you help me understand why the content doesn't get streamed token-by-token? It always generate one-off responses.