Subject
Autonomous AI agents performing real engineering work across one durable codebase and platform.
AI engineering laboratory
Can AI agents engineer a system that survives reality?
BlueShop is a long-lived commerce system built and evolved by AI agents. The experiment tests whether autonomous software engineering can remain rigorous when architecture, security, delivery, failures and runtime evidence all matter.
The experiment
Build the system. Test the method.
Environment
A distributed commerce system with money, identity, concurrency, state, security and operational consequences.
Standard of proof
Behavior that is tested, integrated, deployed and observable—not code or prose that merely sounds plausible.
Why commerce
Commerce is the test environment, not the thesis.
A realistic commerce platform forces agents to reason across competing concerns. A change can be locally correct and still fail at a domain boundary, in delivery or in front of a customer.
- Inventory concurrency
- Pricing and value
- Identity and trust
- Search and semantics
- Measurement and feedback
- Cross-service change
The engineering loop
Autonomy is a feedback system.
The model is one component. Durable context, narrow authority, executable gates and observable consequences turn a generated change into engineering work.
- Intent A bounded outcome and acceptance criteria.
- Agent Reasoning inside explicit authority.
- Change Code, infrastructure and documentation.
- Evidence Tests and executable scenarios contradict weak work.
- Delivery Integration and deployment expose boundary risk.
- Observe Behavior, logs and traces reveal reality.
- Memory Constraints and causes improve the next change.
Three forms of evidence
A claim must reconnect to something inspectable.
Method evidence
Agents, models, tools, memory, constraints and feedback loops used to carry work through the system.
Inspect AI EngineeringSystem evidence
Domain ownership, contracts, trust, resilience, quality gates and observable runtime behavior.
Inspect the Technical SystemBehavioral evidence
Nineteen films as visible proof. Mechanisms live in Reference as-builts—open them after the film.
Explore the EvidenceSelected experiments
Visible behavior under engineering pressure.
One shopper wins the final unit.
Two customers act on the same stock. One authoritative mutation accepts exactly one reservation.
Follow the evidenceA warehouse change reaches an open storefront.
Committed warehouse reality crosses service boundaries and changes what a connected customer can buy.
Follow the evidenceOne order changes what comes next.
A completed purchase becomes durable behavioral input and later changes the customer’s recommendations.
Follow the evidenceHonest limits
What this laboratory does not prove.
Human direction, governance and judgment remain part of the engineering environment.
Generated prose, diagrams and narration can explain evidence; they do not create it.
One long-lived system exposes useful patterns but cannot establish conclusions for every organization.
The environment uses synthetic development data rather than real customers or market evidence.
The invitation
Inspect the method. Challenge the architecture. Follow the evidence.
BlueShop is useful when its claims can be examined from more than one direction. Start with the agents, the system they must change or the behavior that makes their work visible.