feat(samples): add enterprise_dashboard with compact layout patterns

[jswortz]

Summary

• Add enterprise_dashboard sample agent demonstrating compact, visual-first A2UI output patterns
• Include 2 reference A2UI JSON examples (KPI dashboard, store comparison)
• Document anti-wall-of-text prompt engineering patterns in README

Motivation

Every existing sample (restaurant_finder, rizzcharts, personalized_learning) focuses on agent framework integration but doesn't demonstrate how to prevent the LLM from falling back to markdown "walls of text" when presenting data.

Enterprise use cases — dashboards, KPIs, store comparisons, product rankings — are especially prone to this: the LLM defaults to markdown tables, bullet lists, and inline numbers instead of structured A2UI layouts. There's no reference sample showing the right patterns.

What this sample teaches

Layout recipes (data type → A2UI component)

Data Type	Instead of (markdown)	Use (A2UI)
KPI metrics	Inline numbers in paragraphs	Row of Card with bold Text
Comparisons	Markdown table	Row of Card side-by-side
Rankings	Numbered list	List with Card children
Multi-section	Multiple ## headers	Tabs component
Separators	--- horizontal rule	Divider component
Status indicators	Text "up"/"down"	Icon (trending_up, trending_down)

Prompt engineering patterns

1. Anti-markdown rules — Explicit bans with A2UI alternatives
2. Layout recipes — Map each data type to a component composition
3. Output ordering — A2UI JSON first, brief text after (max 1-2 sentences)
4. Component diversity — Minimum 3 different component types per response

Files

File	Purpose
agent.py	Minimal LlmAgent runnable with adk web
prompt_builder.py	Compact layout prompt construction with anti-markdown rules
tools.py	Mock analytics tools (KPIs, store comparison, product rankings)
examples/0.8/kpi_dashboard.json	Reference: 4 KPI metric cards in Row layout
examples/0.8/store_comparison.json	Reference: 3 store cards with Icon indicators
README.md	Pattern documentation and usage guide

Running

cd samples/agent/adk
adk web enterprise_dashboard

Example queries:

• "Show me this week's KPIs"
• "Compare store performance"
• "What are the top 5 products?"

Production validation

These patterns have been battle-tested across 3 production agents deployed on Vertex AI Agent Engine (Gemini 3.5 Flash). Before adding anti-markdown rules and layout recipes, ~30-40% of responses fell back to markdown. After, the rate dropped to <5%.

Related

• feat(agent_sdk): add strict_output mode for reliable A2UI-first output #1465 — strict_output mode for generate_system_prompt() (SDK-level enforcement)
• feat(agent_sdk): add A2UIOutputMode enum for unified prompt generation #1466 — A2UIOutputMode enum for unified TEXT/TOOL prompt generation

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

[gemini-code-assist]

Code Review

This pull request introduces an Enterprise Dashboard sample for the A2UI agent framework, featuring an agent that prioritizes visual component layouts over standard markdown. The implementation includes mock analytics tools, example JSON payloads, and a specialized prompt builder. Review feedback focuses on improving the prompt builder's path handling for better portability, reducing redundancy in system instructions, and ensuring data consistency in mock tool responses. There is also a suggestion to better align example JSON files with the prompt's best practices regarding icon usage.

[gemini-code-assist]

The hardcoded relative path examples/0.8 for examples_path is brittle as it depends on the current working directory. This can cause issues when running the script from different locations, including the if __name__ == '__main__' block. It's more robust to construct an absolute path based on the current file's location.

You'll also need to add import os at the top of the file.

Suggested change

	examples_path="examples/0.8",
	examples_path=os.path.join(os.path.dirname(os.path.abspath(__file__)), "examples/0.8"),

[gemini-code-assist]

This sentence, "You are a VISUAL DASHBOARD, not a text chatbot.", repeats the core identity already established in ROLE_DESCRIPTION. To make the prompt more concise, you could consider removing this line and starting the UI_DESCRIPTION directly with the "Layout Recipes" section.

[gemini-code-assist]

This rule against using markdown tables is also stated more explicitly in the 'Output Rules' section on line 56. To keep the prompt concise and avoid redundancy, consider removing this line. The instruction on line 44 already provides the positive alternative.

[gemini-code-assist]

The transactions field is an integer, while other numeric-like fields such as revenue and avg_basket are pre-formatted strings. For consistency within this tool's return value and to make the data structure more predictable for the LLM, consider formatting transactions as a string as well. This applies to all stores in the list.

Suggested change

	"transactions": 5100,
	"transactions": "5,100",

[gemini-code-assist]

Similar to get_store_comparison, the units_sold field is an integer while revenue is a pre-formatted string. To improve consistency across the tools in this file, consider formatting units_sold as a string with comma separators. This should be applied to all products in the list.

Suggested change

	"units_sold": 1240,
	"units_sold": "1,240",

[gemini-code-assist]

The prompt in prompt_builder.py encourages using Icon components for status indicators. This metric has a downward trend (-1.2%), which is a perfect opportunity to demonstrate the use of a trending_down icon, as suggested in the prompt's Status indicators recipe. The current implementation only uses text. Using an icon here would make the example more consistent with the prompt's best practices and the store_comparison.json example.