NeFut Logo NeFut
Admin Login

[CS.AI] Do Not Restart: Residual Completion for Stateful Agent Handoffs

Published at: 2026-09-15 22:00 Last updated: 2026-09-16 00:22
#Machine Learning #LLM #Artificial Intelligence

Routing and cascades lower tool‑agent cost by transferring control across models, yet stateful handoffs must preserve accepted choices, realized effects, and unfinished obligations. We formalize this as commitment‑constrained residual completion and propose Commitment‑Frontier Residual Completion (CFRC). CFRC enforces target‑before‑proposal, whole‑proposal‑before‑authority, and live‑evidence‑before‑success: it freezes a residual contract from accepted progress, closes the successor continuation into an evidence‑linked graph, and permits execution only when the remainder is covered and live receipts discharge obligations. We establish contract‑relative partial correctness and extend it to the original residual request under complete contract construction.

Across five environments and two same‑provider model pairs, CFRC attains macro accuracy comparable to strong full‑task agents while using only 22.0%‑34.6% of their inference cost. Additional cross‑provider experiments demonstrate broader transfer capability.

Review

Original Source: https://arxiv.org/abs/2609.13800

[h] Back to Home