What moves into software
| Part | What it holds | What the AI agent no longer has to reinvent |
|---|---|---|
| Schema | Fields, units and allowed values | The shape of the form |
| Checks | Required sources and missing-field checks | Whether a required field exists |
| State | Draft, review, released and closed | Where the work stands |
| Output | A repeatable report layout | Formatting on every run |
| Store | One versioned home for each record | Which scattered file to use |
A person chooses the forms and checks. Code is not a substitute for reviewing the answer.
The first build
The first build links FMEA and a control plan on an engineering project. The same records-and-checks pattern is meant for the rest of the business: a customer case, an inventory count, a purchase order or a month-end close.
We're starting with AI agents that keep working and remembering between conversations, because that is the behavior this build is meant to support.
The record store connects through MCP, a standard way for an AI agent to use outside software. People decide what the program does and check it. The AI agents doing daily work use those connections; they do not get permission to rewrite the program.
What the field has found
These are reports from other authors, not Advance Now performance results. The numerical finding belongs to the cited research, not to this build.
| What the field found | Source | The older tool that already does it |
|---|---|---|
| Add a check between steps, so a failed step stops the chain. | “Building Effective Agents” (2024) | Stage gates; ladder logic |
| Constrain the input: fixed choices, allowed ranges, required steps. | Agent poka-yoke write-ups; Thoughtworks Radar | Poka-yoke |
| Keep notes in files outside the conversation, and read them back. | Guide to context engineering (2025) | Travelers, setup sheets, batch records |
| Spread a task over a conversation and performance drops an average of 39%. | Laban and others (ICLR 2026) | State programming: one mode at a time |
| Have a person approve actions that are sensitive or can't be undone. | A practical guide to building agents (2025) | Interlocks; sign-off before release |
| Run the same task many times, and count it as reliable only if it passes every time. | Sierra, τ-bench (ICLR 2025) | Commissioning and capability studies |
Full citations are in the Notes. Structure sets the size of the job compares the main ways to build an AI agent.
What you can try today
List the decisions your AI agent repeats that never change. Put one into a fixed template or a script. Check its output independently. Keep the prompt focused on what requires judgment. Record failures and compare against the previous version before giving it more responsibility. A well-shaped record can still be wrong, so the responsible person checks the evidence and decides.
Public experiments and references
A.N. Creation Office keeps the public AI agents and examples. hotbotspot documents business tools and illustrative records.