Safe, persistent, and evolving agent harness for partially observable worlds
Large language model (LLM) agents can invoke tools fluently, yet enterprise workflows demand stricter safety and state persistence. Tool feedback often hides side effects under partial observability, and long‑horizon tasks require state tracking across multiple records. To meet these needs we introduce E‑Ledger, a multi‑agent harness for safe execution. E‑Ledger adds a code‑approval layer that checks each proposed action against organizational policy before execution and maintains a world ledger that stores verified hidden rules together with evidence‑backed dynamic state.
Because hidden rules are usually unknown beforehand, we propose WorldAbduct, a world‑model‑driven abductive evolution framework. WorldAbduct diagnoses execution trajectories from four complementary views: state consistency, world‑observation gap, policy‑gate correctness, and goal judgment. It hypothesizes latent rules from these views, validates them through targeted abductive interactions, and then integrates the new rules into the ledger.
On the enterprise benchmark World of Workflows, E‑Ledger with WorldAbduct improves safe task completion by 5–15 % across four LLM backbones. Experiments in ScienceWorld and DiscoveryWorld show the same abductive evolution transfers to scientific settings. Code is available at https://github.com/HKUST-KnowComp/E-LEDGER-WorldAbduct
Review: E‑Ledger tightly couples policy auditing, persistent state, and abductive learning, offering enterprise‑grade LLM agents a verifiable and safe execution path, especially effective in partially observable scenarios where hidden rule discovery and adaptation are critical.