The problem
An agent action still needs a human decision.
Cloud automation is useful only when an operator can understand and control what an agent is about to change. The tool boundary needed to combine conversational workflows with the checks expected from infrastructure software.
What I owned
The parts I was responsible for.
I built Model Context Protocol tools and the security controls around them, then connected those tools to cloud and cluster APIs.
- MCP tools for AWS, Kubernetes, and GCP operations.
- Cryptographic authentication and explicit approval for actions.
- Audit records for requested and completed operations.
- Deployment automation, Kubernetes manifests, RBAC, network policy, services, firewall rules, and QR-based TOTP.
Engineering decisions
Separate intent from execution
The agent can prepare an operation, but the operator sees the proposed change and approves it before the tool executes it.
Put security at the tool boundary
Authentication, authorization, and audit behavior travel with the operation rather than depending on the surrounding chat interface.
Return concrete infrastructure state
Tools report the resources they inspected or changed so the operator can verify the result instead of relying on an assistant summary.
Validation
What the implementation had to prove.
- AWS and Kubernetes server deployment flows.
- Kubernetes RBAC, network policy, service, and manifest operations.
- GCP firewall operations and QR-based TOTP authentication.
Claim boundary: Archimedes was proprietary. This page describes the approved public product concept and my technical scope, not private source code, customers, or adoption.