Skip to content

LongGuard 🛡️

The Chain-of-Thought Circuit Breaker for LangGraph and LangChain Agents.

PyPI version PyPI Downloads Python 3.10+ License: MIT Tests: 255 passing Coverage: 90%

What is LongGuard?

LongGuard is a lightweight runtime safety net for autonomous LLM agents.

When agents encounter ambiguous tool outputs or unexpected failures, they often get stuck: - Calling the exact same tool with the exact same inputs 25 times in a row. - Cycling between two reasoning steps in an infinite loop. - Pursuing a dead-end direction that burns 30,000+ tokens.

Frameworks like LangGraph have a built-in recursion_limit, but it is a hard crash: it burns tokens until the limit, crashes with an unhandled GraphRecursionError, and loses the conversation state.

LongGuard catches loops early (typically within 4 steps), injects a "Reflect & Pivot" prompt to help the agent course-correct, and only kills the run gracefully if recovery fails.

CLOSED ──(loop detected)──▶ REFLECTING ──(persists)──▶ HALF_OPEN ──▶ OPEN (Kill)
   ▲                             │                         │
   └──────────(clean step)───────┴─────────────────────────┘

Core Value Proposition

What Happens Without LongGuard What Happens With LongGuard
Agent repeats same tool 25 times Trapped at step 4
Crashes with unhandled 500 error Graceful exit with AI apology message
Burns 40k+ tokens ($2.00–$5.00+ / query) Hard spend cap & early intervention
Agent never knows why it failed Injected Reflect & Pivot prompt forces recovery
Zero post-run insight Detailed GuardReport with step-by-step audit

30-Second Example (LangGraph)

LongGuard wraps your agent nodes in a single line of code:

from langgraph.graph import StateGraph
from longguard.integrations.langgraph import add_guard_to_graph
from longguard import GuardConfig

# 1. Build your StateGraph as usual
workflow = StateGraph(AgentState)
workflow.add_node("agent", call_model)
workflow.add_node("tools", tool_node)
workflow.add_edge("agent", "tools")
...

# 2. Add LongGuard — one line!
workflow = add_guard_to_graph(workflow, GuardConfig())
app = workflow.compile()

Next Steps