BMO — Agent Authority and Verified Software Change

Testing whether bounded AI agents can safely coordinate complex software changes while preserving intended outcomes and independent evidence.

Project Details

BMO is an open-source system developed by Jonathan Maye-Hobbs and JMH Blockchain LLC for coordinating bounded agent work, maintaining decision and authority context, and checking whether the resulting software state meets its intended obligations. This project will test BMO end to end on a scaled API migration in an enterprise-like Azure environment: changes to a service contract, dependent services and clients must preserve authorization, compatibility and other defined requirements. We will compare the final integrated state against independent checks and retain evidence of failures and unresolved outcomes. The current implementation and research are available publicly; the proposed scaled Azure study has not yet been performed.

Theory of Impact

Advanced AI agents can make changes across many connected systems faster than people can manually inspect every interaction. If an agent exceeds its authority, breaks a security obligation, or presents a local success as proof that the entire system is safe, the impact can propagate across dependent services. We aim to evaluate whether explicit authority boundaries, retained decisions, and independent checks on the combined result can expose such failures before changes are accepted. An enterprise-like migration provides a concrete test of these controls under scale and partial failure. Successful controlled results could inform safer oversight for more capable agents; the study would not establish that BMO prevents catastrophic outcomes in general.

People

Jonathan Maye-Hobbs

Team Member

Grants Received– no grants recorded

Add grant

Funding Asks

Funding ask

Open

Minimum

$10,000

Ideal

$30,000

How the money will be spent

Provisional Azure-only budget for a three-month controlled study; no labor costs included. Minimum $10,000: virtual machines and sandbox compute $2,000; Azure-hosted model inference $2,000; storage, databases and network $2,500; monitoring and security services $1,500; contingency $2,000. Ideal $30,000: larger compute fleet $7,600; model inference $7,000; storage and databases $5,500; networking, security and monitoring $4,900; contingency $5,000. The minimum assumes roughly four D4s v5 Linux VMs for three months and 250 million input plus 150 million output tokens; the ideal assumes roughly eight D8s v5 VMs for three months and 1 billion input plus 500 million output tokens. Model-cost examples use GPT-6 Sol Global Standard short-context list rates; estimates depend on region, usage and actual workload.

Discussion

No comments yet. Be the first to share your thoughts.