Building useful agents and showing the evidence: architecture, tools, safeguards, deployment and limitations of systems designed with Claude and OpenAI.
Independent lab. Manda is not affiliated with, certified by, or sponsored by Anthropic or OpenAI.
Claude Code is used as an engineering environment on real products: repository understanding, implementation, review, tests and Git-based delivery. MCP and n8n extend that work through bounded tools and human approval.
The next public track applies the same discipline to OpenAI: typed tools, structured outputs, traces, evaluation sets and action control. The benchmark will publish failures, cost and latency — not only a happy-path demo.
Publishing standard
Every new lab study must make these six elements verifiable.
A bounded business task with an observable outcome.
Model, data, tools, permissions and stop conditions.
Normal, incomplete, conflicting and adversarial cases.
Sensitive actions stay in draft or require approval.
Logs, cost, latency, recovery and model changes.
Known failures and decisions that must remain human.
Available evidence
Product built with Claude Code
An invoicing application connected to business automation: assisted design, Next.js integration, controls and continuous delivery.
Documented MCP experiment
A CMS operated through two MCP servers, covering publishing, SEO and explicitly documented API limitations.
Multi-provider architecture
A Rust agent infrastructure compatible with several providers, including OpenAI and Anthropic, designed for multiple tools and channels.
Next protocol
The next benchmark will compare Claude and OpenAI on a bounded research agent. Planned measures include task success, source quality, tool calls, cost, latency and human recovery.
I can walk through the architecture, code and trade-offs behind these systems — in English or French, remotely for teams in Europe and the United States.
Discuss the system
MANDA