Design the action boundary before designing the agent.
Good agent systems know which tools are available, what permissions apply and when a human must approve.
Independent perspectivesThe publication
The Marmation Review.02 / Engineering
From working prototype to dependable system
Architecture, access control, evaluation, agents, deployment and production integration for technically demanding teams.
Find your next argumentRead across the trust boundary
Use these questions to turn an argument into a reviewable decision.
01Read
Can the wrong identity retrieve the right-looking answer?
Evidence to inspect: Permission and source-revocation tests.
02Decide
Can weak evidence become an authoritative result?
Evidence to inspect: Evaluation cases and no-answer behavior.
03Act
What happens after an approval changes or a request repeats?
Evidence to inspect: State transitions, retry and reconciliation.
An editorial framework. It is not evidence of a completed client system.
Selected reading
10 perspectives / Engineering
Engineering autonomy is a control design problem.
Tool access, identity, approval, idempotency and rollback determine whether an AI agent is useful in a consequential business workflow.
Fine-tuning, RAG or prompting: ask what needs to change.
Model behavior, current knowledge and deployment control belong to different decisions.
Human approval is a system design decision, not a concession.
The pressure to remove people from an AI workflow can obscure the real goal: reliable, accountable work. Sometimes the highest-value automation is one that knows when to stop.
The case for, and against, model-provider abstraction.
Flexibility is valuable when it protects a real dependency. Unnecessary abstraction can hide product and cost differences.
Permission-aware RAG is an identity project
Enterprise retrieval should never turn a document permission into an optional filter.
A RAG evaluation dataset needs more than questions and answers.
A knowledge system must retrieve evidence safely, correctly and consistently as sources change.
RAG is not a custom-trained model. That distinction can save a project.
Retrieval, fine-tuning and model hosting solve different problems. Confusing them can turn a useful enterprise AI initiative into an unnecessarily expensive technical program.
Threat-model the tools, not just the prompts.
An AI agent with business-system access inherits the risk of the actions it can take.
What production-ready AI actually means.
The difference between an AI proof of concept and a production system is not the model's intelligence. It is the system's ability to behave responsibly when conditions are less than ideal.
The editorial approach
Trace identity, evidence and authority through the system. Make its failure behavior as explicit as its capabilities.
AI engineering is often defined by the decisions that never appear in a demo. An agent must be prevented from taking the wrong action. Retrieval must obey permissions. Model versions must be evaluated. An integration must reconcile a successful transaction even after a timeout.
How to use this collection
The Engineering collection studies those systems concerns in concrete terms, including where a packaged platform is better than a bespoke implementation. We prefer readable diagrams, explicit trust boundaries and negative tests over ungrounded claims of autonomy.
Use the articles as thinking tools, not as product advertisements. Start with the question nearest your current decision, follow the cited sources and inspect the limitations. Each article should tell you what to do differently, when the recommendation changes, and which risks cannot be resolved by reading alone.
Talk to an engineer, not a form letter
If the decision leads to a real engineering assignment, share the relevant requirements and constraints. Marmation can help determine whether a focused engagement is appropriate. Discuss a project.
Bring us the
difficult part.
A product idea, a broken workflow, scattered company knowledge or an integration others keep working around. Tell us what needs to change and what has already been tried.
Discuss a project