Overview of Agentic Apps

Agentic apps help in combining your agents to form active, intelligent apps, enabling their use as your decision-making partners. You can consider the agents as workers and the agentic app is the productized solution that bundles those workers with the integrations, knowledge, and controls that are needed. The apps can include multiple agents, each of which is purpose-built for a specific category or domain.

Instead of asking one agent about everything from sales performance to inventory management, you can build specialized agents, for example, a Sales agent, an Inventory agent, a Finance agent; each with deep expertise in their domain. These agents can be used to create your agentic app. Doing so provides better insights and actions because each agent can maintain focused context, domain-specific reasoning, and tailored prompting for its area of responsibility. You can also scale it up by adding additional agents as needed.

The best agentic apps don't simply observe and report; they analyze, prioritize, and recommend actions that drive business forward.

• Proactive alerting over passive monitoring: Agents identify anomalies, opportunities, and risks in real-time, surfacing them with context and urgency. Instead of making users hunt for problems, the agents bring critical issues directly to their attention when they matter.
• Decision support, not data dumps: Apps filter signals from noise by analyzing situations and presenting only actionable decisions. Instead of providing users with every data point, provides synthesized insights with clear recommendations and trade offs.
• Action-oriented intelligence: Every agent interaction leads toward concrete actions, such as, approve a proposal, adjust a strategy, resolve a blocker. If it's purely informational, we can use an API or dashboard instead of agentic apps.
• Context-aware prioritization: Apps understand business goals and user workflows to determine what's urgent versus what can wait. Not all alerts are equal; agents should respect user's time and attention.

Here are the main pillars of agentic apps:

• Information Displays

  Shows you data. These are agent-generated visualizations, such as lists, charts, tables, that show relevant data in real-time using predefined patterns. Agents built for agentic apps answer questions by creating information displays, except when intent clarification, confirmation, or a brief response is more appropriate. Each information display is generated with a topic key that captures the context of the user request. As conversation topics shift, the page template may hide or remove displays that are no longer relevant.

  Agents must be configured to support specific information display types, granting them the capability to generate the displays. The implementation details are provided using prompting. At application startup, each agent may generate an initial information display that's shown to the user immediately. The agent determines whether to show an initial display and which display to show, and the host application can override either decision through prompting.

• Advisors
  Answers your questions. Advisors are agents that answer user questions either independently or in coordination with other agents.

  • Single agent mode: Users focus on a specific agent as their advisor
  • Multi agent mode: Questions are routed to two or more capable agents, and outputs are consolidated into a unified response

  Agents typically respond with both answers and supporting information displays. By default, information displays prioritize showing detailed data within the visualization itself, with accompanying text summarizing key insights. When a single agent acts as advisor, only that single agent produces an information display. In multi-agent scenarios, only selected agents generate displays, with the final answer summarizing their collective output.

• Actions
  Helps you make decisions. Actions are agent-generated decision points that require human approval. Here are the artifacts in actions:

  • Title describing the topic
  • Description of what the action performs
  • Instructions for the agent to execute if the action is called (including commands and parameters)

  This configuration is agent-generated and not modifiable by users. Actions are generally associated with a single agent, and when the action is called, that same agent processes the action. The agent either performs the action immediately or engages the user if further clarification is needed.

  Agents must be explicitly configured to generate actions. The generation logic and invocation handling are defined in the agent's prompting. At application startup, each agent can produce an initial list of priority actions that represent immediate decision points an app user may make. When answering questions, agents can include one or more actions in their response, which are then presented to the user.

  Information displays may also expose actions directly within their pattern. They work identically and their presentation is defined by the associated pattern.

• Communications

  Helps you reach out. Communications are agent-suggested outbound messages, for example, emails or Slack messages that follow consistent predefined templates. Here are the fields in Communications:

  • Title and description explaining what the communication does and who it targets
  • Template parameters used to compose the message

  Communications always initiate outbound messages; they never intercept or receive incoming messages. The communication suggestions are routed to a specific agent for delivery. This agent doesn't need to be an advisory agent. It can be a generic, reusable agent shared across applications. The designated agent then processes the communication request, performs any additional work, and delivers the message using the appropriate tool, such as email, Slack, and Microsoft Teams.

  All communications adhere to the templated patterns for consistency and uniformity. Messages aren't free-form; they conform to specific templates with defined parameter sets. Agents can produce communication suggestions when responding to inquiries and at app initialization. They can prefill template parameters based on their knowledge and prompting. Communication suggestions can also be app-configured and associated with specific agents or all available agents.

  When a communication item is selected, the associated agents attempt to populate any unfilled template parameters. The app user can then modify these values before the message is sent to the template engine for final processing.