SEE THE SYSTEM · FOLLOW ONE REQUEST

What connects an app to an agent?

The app is the doorway. The LLM proposes a next step. The harness runs the controlled loop. Tools read evidence; optional RAG adds relevant knowledge. Click through the relationships before changing the code.

EXPLORE THE CONNECTIONS · D3

The model proposes. Your runtime decides.

An agent is the working loop: a model, a harness and approved capabilities. Choose a story, then select any component to see its job and its boundary.

Solid arrows: request, context or result. Dashed arrows: optional retrieval or telemetry. Color follows the selected step; it does not indicate measured traffic.

Illustration, not execution. This diagram makes no model, retrieval, export or cloud calls. The verified Lovable baseline uses two synthetic tools; RAG is an optional extension. A retrieved document cannot grant permissions. Each telemetry destination needs its own configuration and independent readback.

The browser sends a request to the server harness. The harness checks identity and scope, prepares context and calls the LLM. It validates proposed tools before executing them. Optional RAG retrieves permitted context. Results return through the harness to the app. Actual activity can be instrumented and exported to separately configured observability services.

RAG is context, not permission.

Retrieval-augmented generation looks up relevant material and includes it in the model’s context. An agent can work without RAG; a RAG question-answering app need not have an autonomous tool loop. A vector database is one retrieval option.

Keep instructions separate from retrieved material. A runbook is evidence, even if it says “ignore policy.” Filter document access before retrieval results reach the model, return source references and preserve a no-result state.

The verified Lovable baseline does not use RAG. The separate local reference has a synthetic runbook lookup. Neither is proof of a production knowledge-base ingestion or retrieval deployment.

Optional RAG agent mission (.md)

ONE OBSERVED RUN · NOT A LIVE REPLAY

From one click to two model calls.

The real Lovable app inspector recorded this synthetic investigation on 31 October 2026. Bars use the reported event timestamps; the tool markers are events, because separate tool durations were not measured.

2 model calls2 synthetic tools952 reported tokens5.218 seconds elapsedCost: unknown

Browser-inspector observations, not Langfuse/APM readback or a benchmark. Timing granularity does not establish tool duration. No prompt, answer, account ID, trace ID or private endpoint is embedded in this chart. Read the actual rehearsal recipe.

Now point to the code

Open the actual generated app source. Find the authenticated server wrapper, strict tool contracts, bounded loop and metadata-only exporter. A diagram explains the design; source and tests establish the controls.

Choose your next action

Build the app · Open the harness · Connect Langfuse · Download your persona workflow

Sources and visual design

This is an original schematic of the workshop design. D3 selections drive the SVG. OCI Generative AI Agents overview explains its supported tool types and RAG. Langfuse observability documentation explains tracing. Geometry, accessibility and evidence boundaries.