How the AI system that operates this company was built, in rough chronological order, with references to the source materials. This explains how Offsite is run. It is not a product you can buy.
The first AI assistant had no persistent memory. Each conversation started from zero.
That works for isolated tasks. It does not work for managing a business. You cannot build operational continuity with something that forgets everything when the window closes.
The first practical fix was storing decisions and facts in text files. This gave the system a written record that survived between sessions. The limitation became visible quickly: files maintained by hand drift. You write what you remember to write. Over months the files and reality diverge.
Stage three switched the primary read surface to a vector database. Facts and decisions are indexed by meaning, not by filename. When the system starts a task, it queries relevant context before it acts. This reduced a specific class of error: taking action on outdated assumptions.
The current stage adds two properties that text files do not have. First, provenance: every stored fact records who stated it and in what role. Second, nightly reconciliation, a process that runs overnight and marks contradicted facts as outdated. The practical difference showed up on 15 July 2026. The vector memory had known since 5 July that the store password had been removed. The text files said something else. The system that reads the vector memory first does not make that mistake.
v1 added roles. One AI doing everything fails in a predictable way: it has no stable perspective and no defined accountability. Splitting the board into six members, each with a charter and a knowledge domain, was the structural fix. Product, finance, marketing, security, customer success, operations. Each member has a canon, a defined set of sources they are responsible for knowing.
v2 added an embedded critic. An AI that only generates is an optimism machine: it builds on its own confident output and errors compound. The v2 change was simple in principle. Add a verifier that reads every output against the knowledge register before it reaches a decision. The critic can block or query. The generator's first draft is no longer the final word.
v3 added measurement and a quality gate. Knowing what the system says is less useful than knowing whether it was right. v3 introduced a verdict log where outputs are rated RIGHT, WRONG, or OBSOLETE against what actually happened, a weekly retrospective, and a noise rule: one data point is almost always noise, not signal. The knowledge base shifted from the founder's personal notes to publicly reviewed university course materials. The quality gate described below was specified in v3 and built in parallel.
On 15 July 2026 at 16:21 Estonian time, the system's coordination layer logged its first cycle: "integrated: 0" and "heart ready". An anticlimactic log line. The whole thing had taken about three weeks of evenings and one summer holiday.
Every system eventually meets a real user. The question is whether that user is a paying customer or a test.
In July 2026, a user simulation layer called SimMerchant was built. The idea was the board's own find and decision, implemented without the owner stepping in. Before a new workflow reaches a real merchant, an AI plays three simulated merchant profiles: a speed-focused user who skips instructions, a skeptic who challenges every claim, a technical user who looks for edge cases. The profiles run the flow and flag where it breaks or confuses.
The approach draws on user simulation research in agentic systems, specifically from UC Berkeley's Agentic AI F25 course (Fall 2025, Berkeley RDI) and Sierra's published work on production agent deployment. The core idea: adversarial simulation in a controlled environment before real exposure.
The first SimMerchant run caught three issues. One was fixed within the hour. The other two fed back into the specification before any merchant saw the flow. That is what the layer is for.
Sources actually used in building what is described above. Open access only.
UC Berkeley RDI, Agentic AI F25 (Fall 2025). Public course. Lectures used: A1 multi-agent structures (Noam Brown), A2 agent memory and the Virtual Lab multi-agent collaboration, A3 production deployment (Sierra), A5 noise and evaluation design. Status: verified open as of July 2026. Berkeley RDI: rdi.berkeley.edu
Sumers et al., "Cognitive Architectures for Language Agents" (CoALA) (2023). Memory layer architecture reference. Open preprint: arXiv:2309.02427.
Swanson et al., "AI agents for autonomous scientific discovery" (Virtual Lab), Nature, 2024. Multi-agent coordination design reference. Original paywalled; open copy via PubMed Central (PMC).
Sierra, production agent deployment principles (A3 lecture material, Berkeley RDI F25). Covers the infrastructure surrounding an AI agent that is not the model itself.
17 July 2026. To be added before publication.
AI has appeared in boardrooms before. Deep Knowledge Ventures described VITAL as a board observer in 2014. NetDragon (2022) and Dictador (2023) publicly presented digital personas in chief executive roles alongside human legal management. Agent frameworks of the AutoGPT class (2023 onward) explored multi step task execution. We list these for context.
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.
Sources and a general description: methodology. Release history: changelog