Platform architecture principles.

Every agent must define explicit human oversight checkpoints. Irreversible actions - sending emails, writing records, triggering workflows - require an approval gate before execution. Autonomous action without oversight is not acceptable on high-impact paths.

Agents must not retain customer data between sessions. All state is explicit, documented, and held in a named store. Implicit memory across turns is a privacy risk and a compliance failure. Architects must trace every data write at design time.

OTel instrumentation, structured logs, and an evaluation suite are required before any production deployment. If you can't observe it, you can't run it. Monitoring dashboards and alert thresholds are part of the definition of done, not an ops afterthought.

Prompt injection, insecure output handling, data leakage, over-permissioned tools, and supply chain vulnerabilities are treated as blockers. A threat model covering LLM-specific attack surfaces is required before code is written, not reviewed at launch.

ISO 42001 controls applied. Bias, fairness, and transparency statements documented before launch. Agents processing HR, health, or financial data require a fairness test plan covering protected characteristics. Results are logged and reviewed on a defined cadence.

Domain knowledge comes from RAG and grounding, not model fine-tuning. This keeps knowledge auditable, updatable, and version-controlled. Fine-tuning is a last resort - it makes the knowledge opaque, expensive to update, and difficult to audit for correctness.

OneTrust intake required before any personal data is processed. Every personal data flow documented in a RoPA entry. Lawful basis confirmed. Data minimisation applied - agents request only the data required for the task, not the data available to the tool.

Automated evaluation pipelines run on every merge. A deploy requires a passing evaluation score, not just passing unit tests. Evaluation covers accuracy, safety, tone, and refusal behaviour. A score below threshold is treated the same as a failing build.

Every agent ships with a runbook and a monitoring dashboard. If you can't explain how to operate it without you in the room, you haven't finished building it. Operational documentation is part of the acceptance criteria, reviewed at the same time as the code.