The scientific and governance sources behind the engine, at the level we can publish. This describes how the company is run, not a product for sale.
A management layer of six persistent AI roles over a shared, searchable memory. Every significant output passes a critic step before it becomes a decision. Decisions are logged and later graded against what actually happened. Actions that touch production run through defined quality gates, and significant records are kept in a form that shows if they are changed later. That is the whole picture at the level we publish.
UC Berkeley RDI, Agentic AI course, F25 (Fall 2025). Public course. Used for multi-agent structures, agent memory, production deployment and evaluation design. rdi.berkeley.edu
Sumers, Yao, Narasimhan, Griffiths. Cognitive Architectures for Language Agents (CoALA), 2023. Open preprint, arXiv:2309.02427. Memory layer reference.
Swanson et al. AI agents for autonomous scientific discovery (the Virtual Lab), Nature, 2024. Multi-agent coordination reference; open copy via PubMed Central.
Sierra, production agent deployment principles, via the Berkeley RDI course. The infrastructure around an agent that is not the model itself.
Anthropic, Building Effective Agents, 2024. Simple, auditable agent designs and limited tool authority.
OpenAI, Practices for Governing Agentic AI Systems, 2023. Human intervention points, attribution, narrow default authority.
NIST AI Risk Management Framework 1.0 (2023) and the Generative AI Profile (2024). Audit trails, defined authority, incident handling.
G20/OECD Principles of Corporate Governance, 2023. The split between a supervisory board and a management board that our governance model borrows.
The board is modelled on how real management boards work, not on chat assistants. The sources we studied: the G20/OECD Principles of Corporate Governance (2023) for the split between a supervisory board and a management board; the Cadbury Report (1992) for the practice of a schedule of reserved matters; the Estonian Commercial Code on management board duties and the duty of care (Offsite itself is registered in Estonia, in the European Union, so these are the rules the company operates under); the Estonian Corporate Governance Code with its comply or explain principle; the IIA Three Lines Model (2020) for independent internal audit; the audit independence and business judgment tradition in company law; and Stanford GSB governance research (Larcker and Tayan, Corporate Governance Matters).
Open agent frameworks, Hermes among them, are engines: they run tasks. What we are building is a different layer: an AI management board for a real, registered legal entity, operating under the governance model described above.
The responsibility model is the clearest difference. In a real company the management board carries a duty of care and answers for its decisions. Here, legal responsibility stays with the owner and the legal entity. When the board makes a wrong call, the consequence is reconfiguration: a root cause analysis, an engine update approved by the owner, and a public log entry. The system is corrected rather than punished.
Memory is the clearest technical difference. Recall based memory in most frameworks appends what it sees and retrieves what looks similar. Ours is curated toward current truth: every stored fact carries its source and role, external input is quarantined before it can become operating truth, a nightly process reconciles stored facts against live systems, and superseded facts are removed rather than accumulated. Unrelated business contexts are kept strictly separate at the data and credential level.
We adopt good ideas from the open ecosystem where they fit, and we say so. Recent examples: an automatic step that asks after every significant run whether a lesson should be recorded, and an evaluation of open, community shared skill standards. The engines themselves are becoming a commodity. Our work sits in the governance, memory and accountability layer on top.
Architecture details, internal process mechanics, timing of automated operations, and anything that would allow a direct replication of the system. The test before any disclosure: could a competitor build the same thing from this information? If yes, it does not go out.
Back to the engine story.