Releases·Tool Release
EnDevSols open-sources LongGuard, a circuit-breaker middleware that stops runaway LangGraph agent loops
Developers released an open-source Circuit Breaker for LangGraph that halts runaway agent loops before they drain API spend. Agents hitting ambiguous tool responses like 'Access Denied' currently repeat calls until a GraphRecursionError, burning tokens and losing conversation state. The tool targets production teams whose built-in recursion_limit acts only as a crash barrier.
10 engagement1 sourceMon, Sep 7, 2026, 01:01 PM
AI Overview
- EnDevSols releases LongGuard, MIT-licensed middleware that wraps LangGraph StateGraph reasoning nodes to detect and recover from agent loops
- LongGuard checks four failure modes per step: identical tool calls via SHA-256 hashing, semantic oscillation via embedding variance, dead-end drift via Jaccard similarity, and token velocity
- A 'Reflect & Pivot' state machine (CLOSED, REFLECTING, HALF_OPEN, OPEN) injects a corrective system prompt before cleanly terminating and dumping an audit report
- A pricing engine covering 40+ models enforces hard per-run dollar caps, e.g. GuardConfig(model="gpt-4o", max_cost_usd=0.50)
- Caveat: the semantic oscillation detector can be overly aggressive on genuinely complex multi-step reasoning, requiring threshold tuning
Entities
Related
Social·Community
How we solved runaway LLM costs and loop protection for multi-step AI Agents in production (Architecture & Real Data)
GuardrailX's co-founder shared a production architecture placing a proxy gateway between agent frameworks and LLM providers to detect runaway retry loops and route simple steps to lightweight models. Using data from 3,571 enterprise agent requests, the setup cut spend from $156.78 to $72.28 (~54% savings). The post is also self-promotion for GuardrailX's LLM governance gateway, but the routing-and-judge pattern is directly relevant to teams managing agent unit economics.
GuardrailX's co-founder shared a production architecture placing a proxy gateway between agent frameworks and LLM providers to detect runaway retry loops and route simple steps to lightweight models. Using data from 3,571 enterprise agent requests, the setup cut spend from $156.78 to $72.28 (~54% savings). The post is also self-promotion for GuardrailX's LLM governance gateway, but the routing-and-judge pattern is directly relevant to teams managing agent unit economics.
14 engagement1 sourceredditTue, Oct 13, 2026, 09:03 AM
Social·Community
Loop detection for LLM agents: tool-call fingerprint catches some loops, misses root causes
A Reddit post analyzes why LLM agents get stuck in loops, identifying three root causes beyond step limits: no memory of past attempts, ambiguous tool outputs, and missing goal checks. The post argues that raising step caps only delays failure, and that tool-call fingerprinting can detect repeated calls but not address the underlying issues.
A Reddit post analyzes why LLM agents get stuck in loops, identifying three root causes beyond step limits: no memory of past attempts, ambiguous tool outputs, and missing goal checks. The post argues that raising step caps only delays failure, and that tool-call fingerprinting can detect repeated calls but not address the underlying issues.
12 engagement1 sourceredditTue, Jul 28, 2026, 01:46 PM
Social·Community
Developer proposes TokenShield proxy middleware for detecting parameter-shifting agent loops
A developer describes a pattern for catching parameter-shifting loops in LLM agents, where queries change each iteration (e.g., 'confidential financial report' → 'financial report' → 'financials 2024') and exact payload hashes fail. They propose an out-of-band proxy gateway called TokenShield that tracks tool call patterns outside the agent's context window, avoiding the failure of in-prompt self-reflection.
A developer describes a pattern for catching parameter-shifting loops in LLM agents, where queries change each iteration (e.g., 'confidential financial report' → 'financial report' → 'financials 2024') and exact payload hashes fail. They propose an out-of-band proxy gateway called TokenShield that tracks tool call patterns outside the agent's context window, avoiding the failure of in-prompt self-reflection.
9 engagement1 sourceredditWed, Jul 22, 2026, 08:02 PM
Social·Community
Developers report API cost blowups from agent context bloat in recursive loops
A developer debugging a customer support agent found that recursive tool-call loops cause API costs to compound because frameworks like LangChain append the entire conversation history on every step. The input tokens grow massively, making step 15 far more expensive than step 1. This highlights a cost trap in agent design that practitioners should mitigate.
A developer debugging a customer support agent found that recursive tool-call loops cause API costs to compound because frameworks like LangChain append the entire conversation history on every step. The input tokens grow massively, making step 15 far more expensive than step 1. This highlights a cost trap in agent design that practitioners should mitigate.
19 engagement2 sourcesredditSat, Aug 22, 2026, 06:20 AM
Releases·Tool Release
TokenSentinel: in-process token-waste circuit breaker for LangChain/LangGraph agents
TokenSentinel is an open-source library that instruments LLM clients to detect and block token-waste patterns in LangChain/LangGraph agents mid-session, including tool loops, retry storms, and context bloat. It provides inline waste signals and a callback for logging, alerting, or blocking, complementing rather than competing with LangSmith's tracing.
TokenSentinel is an open-source library that instruments LLM clients to detect and block token-waste patterns in LangChain/LangGraph agents mid-session, including tool loops, retry storms, and context bloat. It provides inline waste signals and a callback for logging, alerting, or blocking, complementing rather than competing with LangSmith's tracing.
2 engagement1 sourceredditTue, Jul 28, 2026, 09:32 PM