Safe AI Agent

AI agent with safety guardrails, evaluation harness, prompt management, and the Platform tenant chart whose ModelGateway is its model path. The responsible AI stack — filter inputs, validate outputs, test everything.

Name	`safe-ai-agent`
Version	`0.1.0`
Tags	ai, safety, guardrails, agent, eval
Source	`composites/safe-ai-agent.yaml`

Render it

# CLI
npx @nanohype/sdk render-composite safe-ai-agent --out ./my-app

# SDK
import { LocalSource, renderComposite } from "@nanohype/sdk";
const result = await renderComposite(source, "safe-ai-agent", variables);

# MCP — from an agent
get_composite({ name: "safe-ai-agent" })

What it draws together

Template	Rendered at	Notes
`monorepo`	root	the root project
`agentic-loop`	`packages/agent`
`guardrails`	`packages/guardrails`
`eval-harness`	`packages/evals`
`prompt-library`	`packages/prompts`
`k8s-app-tenant`	`infra`

Variables

Required

ProjectName string: Project name Must be lowercase kebab-case.

Optional

LlmProvider string — defaults to gateway: LLM provider the packages deployed on the substrate call: gateway, the Platform ModelGateway, which holds the Bedrock role there, so none of them holds a model credential. The eval harness runs on a CI runner and keeps its own bedrock default
Tenant string — defaults to ${ProjectName}: Owning team: the tenants-<team> namespace the Platform, BudgetPolicy and ModelGateway CRs are applied in, the Platform's spec.tenant, and the team label on the tenant chart's resources and telemetry
Image string — defaults to ${ProjectName}: Container image repository the tenant chart deploys, without a tag; each values-<env>.yaml pins one
AppMetric string — defaults to http: Prefix of the RED series the tenant chart's Grafana dashboard reads: <prefix>_requests_total, <prefix>_errors_total and <prefix>_request_duration_seconds. The default, http, is the prefix a ts-service emits them under; set it to the prefix the deployed service emits