S-Bus uses HTTP logs to stop LLM agent state corruption
S-Bus is an HTTP middleware that prevents multi-agent LLM state corruption by reconstructing each agent's read set from HTTP traffic, then blocking conflicting writes — machine-checked in TLAPS and validated across 427,308 conflicts with zero data loss.
S-Bus is an HTTP middleware that prevents structural race conditions when concurrent LLM agents share mutable state. The core mechanism is a server-side DeliveryLog that reconstructs each agent's read set by passively tracking HTTP GET operations, then blocks conflicting writes at commit time — all without requiring changes to agent SDKs or frameworks like LangGraph, CrewAI, or AutoGen.
The system enforces Observable-Read Isolation (ORI), a partial causal consistency property that prevents write-write and stale-read conflicts across shared state shards. Author Sajjad Khan provides three tiers of mechanized proof: ReadSetSoundness and ORICommitSafety machine-checked in TLAPS (the TLA+ Proof System), exhaustive model checking in TLC covering 20,763,484 distinct states at N=3 agents with zero violations, and nine inductive soundness lemmas discharged in Dafny. Empirical tests against PostgreSQL 17 SERIALIZABLE and Redis 7 WATCH/MULTI show conflict-prevention parity — 427,308 active HTTP-409 conflicts handled with zero Type-I data corruptions across all three backends.
One caveat: ORI's operating envelope is topology-conditional. It remains semantically neutral when agents work on dedicated shards, but can propagate concurrent contradictions in single-shard collaborative writing scenarios. Source code is available on GitHub.