langgraph: Instructions file for Codex

AGENTS Instructions

This repository is a monorepo. Each library lives in a subdirectory under libs/.

Corridor security analysis

When Corridor's analyzePlan tool is available, create a plan and use the tool to analyze it before generating or modifying code. Apply the resulting security guidance before writing code.

When you modify code in any library, run the following commands in that library's directory before creating a pull request:

• make format – run code formatters
• make lint – run the linter
• make test – execute the test suite

To run a particular test file or to pass additional pytest options you can specify the TEST variable:

TEST=path/to/test.py make test

Other pytest arguments can also be supplied inside the TEST variable.

Direction: LangGraph is a runtime

LangGraph is consolidating as the low-level runtime for agents and workflows: graph execution, checkpointing, streaming, and interrupt/resume. Critical production systems and agent harnesses run on it, so correctness and stability come first. It is moving away from being a friendly SDK for building agents directly; agent development belongs one layer up, on a harness — Deep Agents (github.com/langchain-ai/deepagents, a separate repo) is the recommended one.

Optimize contributions for runtime correctness, reliability, performance, and simplicity — not for the shortest agent example.

• One way to do each thing. A change never adds a way; removals are always deliberate (see below). The number of ways only goes down. No aliases, overloads, extra accepted types, implicit defaults, or customization hooks for convenience: ergonomics alone never justify new public surface.
• The mandate applies to everything that implements runtime behavior, in whichever repo it lives: the graph engine (langgraph, langgraph-core), the checkpointers (checkpoint*), and the server (langgraph-api). The client SDKs and CLI stay ergonomic but are thin, faithful clients, never a second implementation of runtime behavior.
• Agent-level packages (prebuilt, langgraph-supervisor, langgraph-swarm, langgraph-cua) are maintenance-only: fix bugs, don't add agent features. New examples and docs that build agents should use Deep Agents.
• Remove deliberately, not opportunistically. Removing or changing public behavior must be the stated purpose of the PR and go through the deprecation process — replacement in place, removal version recorded, migration note. Never drop a default or overload as a side effect of an unrelated change.
• Behavior changes land in Python and JS together (the implementations live in separate repos; note in the PR when the other one needs a matching change).
• If a task asks for something this section rules out, say so and propose the runtime-level or harness-level alternative instead of quietly implementing it.