Truth Maintenance with Consequences: Recovering Multi-Agent LLM Work After Post-Commit Invalidation
Year: 2026
Type: preprint
Abstract: A multi-agent workflow can finish with every tool call reporting success and still be wrong, because a fact behind one of its decisions is corrected later. By then other agents have built on the result, and some effects, such as a sent notification, cannot be undone. Durable execution handles a crashed worker and sagas a failed step. Neither handles a step that succeeded on a wrong premise, a problem we call post-commit invalidation. Intrusion-recovery systems repaired a similar situation by rolling back and re-executing the affected requests, but that needs deterministic replay, and a model given corrected input makes a new decision rather than a corrected copy of the old one. We propose CARA (Causal Agent Recovery Architecture), which recovers forward instead. CARA records which decisions, artifacts and external effects depend on which observations, finds the work a corrected premise reaches while the system keeps running, and gives each affected item one disposition: derived state is discarded or withdrawn, reversible effects are compensated, and irreversible effects become residual obligations with an owner. Agents pass each other artifacts, never whole conversations, so work that never received the bad premise is kept. We state seven conditional safety invariants and walk one provisioning scenario through the design. CARA is a design proposal; we have not implemented or measured it.
Author: Keshri, Prasanjeet
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236083547
