What is new in AuraForge Foundry
AuraForge Foundry combines several ideas that are usually found separately, or not at all, in AI software tools. This page explains each one in plain terms. For the components themselves, see the product overview.
Five differences
Deliberation across providers, not one model
The usual approach
A typical AI assistant sends a question to one model and returns that model’s answer.
AuraForge Foundry
AuraForge Foundry asks several frontier models from different providers to propose approaches and then review each other’s proposals. Models trained differently tend to make different mistakes, so comparing them shows weak spots that a single model would miss. The final plan keeps a record of where the models agreed and disagreed.
Governance on every request
The usual approach
AI rules are usually set once, as an account setting or a written policy that people are expected to follow.
AuraForge Foundry
In AuraForge Foundry, the AuraForge Supervisory Engine (AFSE) checks each individual request before it leaves: which provider, what data, what cost. Each check is logged, so an organisation can later see exactly what was sent where and why it was allowed.
Cost-aware routing across models
The usual approach
Using the largest available model for every step is simple but expensive.
AuraForge Foundry
Not every step needs the most capable model. AuraForge Foundry is designed to send routine steps to smaller, cheaper models and keep the most capable models for planning and deliberation, within spending limits that AFSE enforces.
Memory that carries across runs
The usual approach
Most AI sessions start from a blank page and forget what happened once they end.
AuraForge Foundry
The Memory Fabric keeps plans, decisions, reviews and outcomes. Later runs can reuse what worked and avoid repeating what did not.
Isolated execution lanes
The usual approach
AI coding agents often run directly on a developer’s machine, with broad access.
AuraForge Foundry
The AuraForge Execution Engine (AFEE) runs each agent lane in its own container. Lanes are separated from each other and from the host, and each lane’s output is reviewed on its own.
Principles behind the design
- Knowledge Before Code
- Understand the problem, the users and the options before writing software.
- Product Before Technology
- Technology is chosen to serve the product, not the other way round.
- Intelligence Before Automation
- Automate a workflow only after it has been understood and checked.
- Humans Remain Responsible
- AI models assist; people stay accountable for decisions, ethics and data safety.
What we are still working on
AuraForge Foundry is in development. Open questions include how best to combine disagreeing model outputs into one plan, how to keep policy checks fast enough not to slow work down, and what should and should not be kept in long-term memory. We describe the design here as it stands today; it will change as we learn.