AI

We Query, Therefore We Compute: On Oracle Computation beyond the Machine, with an Application to Agents

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 08/10/2026, 01:18:24 UTC.

Briefings

This article proposes a unified computational model for agentic systems by treating Large Language Models as Oracles and extending a stack-based automaton to perform Turing-complete computation. The model defines two distinct operations: an Oracle computation that appends answers to a growing stack, and a Priestess computation that manages the stack's growth and synchronization. These two processes synchronize at the end of every answer to model caching and scheduling, with the boundary between them fixed by the Oracle's content. The architecture is realized through ArchNights, an extended RISC-V ISA and Linux-style OS, which runs on gem5 as a testbed. The work establishes a foundational view where agents are computer systems, enabling future research to share invariants and bounds.

Key points

• The proposed model treats Large Language Models as Oracles that append answers to a growing stack, while a separate Priestess computation manages the stack's growth and synchronization.
• The boundary between Oracle and Priestess computations is fixed by the Oracle's content, ensuring a consistent view of the two forms of agentic system.

Context

Agentic systems currently use disparate abstractions like scheduling and caching borrowed from operating systems, creating a fragmented view of Workflows and Agents. This research constructs a unified abstract machine to provide both mechanisms, offering a foundation for future research to share invariants and bounds.

Limitations

The model applies to internally autoregressive Oracles, but no guarantee holds for all Oracle types. The boundary between Oracle and Priestess computations is fixed by Oracle content, not by Oracle type. The implementation relies on gem5 as a computer system, and the ArchNights ISA is extended rather than standard RISC-V. The study is presented as a preprint with no guarantee of final publication status.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 03/11/2026, 21:48:48 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeico6ve5mpdhiics42ulu5a6yyvcqzsxyu6bxslypxotbymgu3hu2q