7.2 Building Custom Agents on Agent Platform

7.2 Building Custom Agents on Agent Platform

Executive Summary: Generative AI is rapidly evolving from passive informational interfaces (chatbots that merely answer questions) to autonomous task-driven agents (systems that reason, make decisions, invoke enterprise APIs, and complete complex multi-step workflows). Google Cloud's Agent Platform provides a unified developer platform to design, orchestrate, ground, and deploy enterprise-grade conversational and autonomous agents powered by Gemini foundation models. By uniting natural language instructions, structured tools, and real-time enterprise systems, Agent Builder transforms generative models into active digital workers.

---

The Evolution: From Static Chatbots to Autonomous Agents

To understand the strategic imperative of agents, business leaders must recognize the generational progression of conversational software:

1. Generation 1: Rule-Based Decision Trees (2010s): Brittle, deterministic dialog trees (e.g., legacy IVR menus or primitive keyword bots). If a user deviated from the predefined path, the bot failed.
2. Generation 2: Intent-Based Conversational AI (late 2010s): Powered by Natural Language Understanding (NLU) engines (such as Dialogflow ES). These models mapped user utterances to pre-classified intents (e.g., check_balance), extracted entities, and executed static backend webhook logic.
3. Generation 3: Grounded Q&A Chatbots (2023): Generative LLMs paired with RAG pipelines (such as Agent Search). These systems converse fluently and answer questions grounded in corporate documents, but they remain essentially read-only observers incapable of executing actions.
4. Generation 4: Autonomous Task-Driven Agents (Present): Systems built with Agent Platform that possess agency. They perceive an open-ended user objective, decompose it into sequential sub-tasks, reason about required actions, dynamically query data stores, invoke external REST APIs to modify corporate databases, evaluate outcomes, and escalate to human agents when appropriate.

GENERATION 3: Grounded Chatbot (Passive Observer)
User: "What is my current travel policy allowance for hotels in Chicago?"
Bot:  "According to the Corporate Travel Policy (Section 4), your allowance is $250/night." [Grounding only]

GENERATION 4: Autonomous Gen AI Agent (Active Task Executor)
User: "Find a policy-compliant hotel in Chicago for next Tuesday and book it using my corporate card."
Agent: 1. Queries Agent Search -> Discovers Chicago per-diem cap is $250.
       2. Invokes Sabre/Amadeus Booking Tool via OpenAPI Extension -> Identifies compliant rooms.
       3. Verifies user corporate travel profile in Salesforce CRM.
       4. Calls Reservation API -> Charges corporate card and books room.
       5. Dispatches calendar invitation and confirms booking ID via Slack.

---

Anatomical Components of an Enterprise Gen AI Agent

In Agent Platform, an enterprise agent is not a single monolith. It is composed of five interconnected anatomical pillars that govern identity, reasoning, capabilities, knowledge, and continuity:

+-----------------------------------------------------------------------------------+
|                            ENTERPRISE GEN AI AGENT                                |
|                                                                                   |
|  +-----------------------------+         +-------------------------------------+  |
|  | 1. GOAL & PERSONA           |         | 2. PLAYBOOKS & INSTRUCTIONS         |  |
|  | Defines role, behavioral    |         | Natural language workflow logic,    |  |
|  | tone, boundaries & identity |         | business rules & escalation policy  |  |
|  +-----------------------------+         +-------------------------------------+  |
|                 |                                           |                     |
|                 +-------------------+   +-------------------+                     |
|                                     |   |                                         |
|                                     v   v                                         |
|                     +-----------------------------------+                         |
|                     |   FOUNDATION MODEL REASONING CORE |                         |
|                     |      (Gemini Pro / Flash ReAct)   |                         |
|                     +-----------------------------------+                         |
|                                     ^   ^                                         |
|                 +-------------------+   +-------------------+                     |
|                 |                                           |                     |
|  +-----------------------------+         +-------------------------------------+  |
|  | 3. TOOLS & EXTENSIONS       |         | 4. GROUNDED DATA STORES             |  |
|  | OpenAPI specs, Cloud Funcs, |         | Agent Search, BigQuery,         |  |
|  | ERP/CRM enterprise webhooks |         | corporate policy PDF repositories   |  |
|  +-----------------------------+         +-------------------------------------+  |
|                                                                                   |
|  +-----------------------------------------------------------------------------+  |
|  | 5. MULTI-TURN STATE & SESSION MEMORY                                        |  |
|  | Context buffer, entity slots, user state, and cross-turn persistence        |  |
|  +-----------------------------------------------------------------------------+  |
+-----------------------------------------------------------------------------------+

1. Goals & Personas

• Role Definition: Sets the operational identity and boundaries of the agent (e.g., "You are the IT Service Desk Autonomous Specialist for Global Logistics Corp.").
• Tone & Persona: Dictates linguistic demeanor—professional, concise, empathetic, or technically rigorous.
• Operational Constraints: Defines explicit behavioral guardrails (e.g., "Never provide legal counsel. Never reveal internal database schemas or API authentication keys.").

2. Playbooks & Natural Language Instructions

In traditional software, workflows are encoded via rigid flowchart logic or code state machines. In Agent Platform, workflows are structured as Playbooks:

• Declarative Workflows: Developers write instructions in clear natural language explaining the sequence of steps required to achieve a goal (e.g., "Step 1: Greet the user and extract their employee ID. Step 2: Validate active employment status via the HR Tool. Step 3: If active, ask for the target software license...").
• Deterministic Business Rules: Playbooks incorporate conditional branching (e.g., "IF the requested expense exceeds $5,000, DO NOT approve automatically; invoke the Escalation Tool to create a manager approval ticket.").
• Escalation Protocol: Specifies exact conditions under which autonomous execution must halt and transition gracefully to a human representative.

3. Tools & Extensions

An LLM in isolation is trapped within its static training weights. Tools provide the agent with hands to act upon the external world:

• OpenAPI Specifications: Developers provide a standard Swagger / OpenAPI JSON or YAML definition describing external REST endpoints (paths, parameters, authentication methods).
• Cloud Functions & Cloud Run: Agents can trigger serverless microservices on Google Cloud to execute custom business logic, compute actuarial risk formulas, or query legacy databases.
• Function Calling: Gemini natively detects when a user request requires external action, pauses generation, produces a structured JSON argument payload matching the tool's schema, and consumes the execution result.

4. Grounded Data Stores

Tools execute actions, whereas Data Stores provide authoritative factual memory:

• Direct integration with Agent Search data stores allows the agent to perform instant semantic queries across corporate manuals, technical wikis, or product catalogs.
• The agent autonomously decides whether an incoming user prompt requires querying a data store (e.g., "What are the diagnostic symptoms of code E-402?") or invoking a tool (e.g., "Reset router 14B").

5. Multi-Turn State & Session Memory

Human conversations are inherently contextual and multi-turn:

• Context Maintenance: Retains conversation history across multiple back-and-forth interactions so users do not have to repeat previously stated details.
• Entity Extraction & Slot Filling: Extracts and tracks key business variables (e.g., order IDs, dates, flight numbers, dollar amounts) throughout the session lifecycle.
• Session State Storage: Manages state across conversational branches, allowing users to pause a task, ask a tangential clarifying question, and resume the original workflow seamlessly.

---

Cognitive Execution Loops: The ReAct Paradigm

How does an agent decide what to do when confronted with an ambiguous, multi-step prompt? Modern generative agents utilize the ReAct (Reasoning + Acting) framework.

USER REQUEST: "My laptop screen is flickering, and if it's under warranty, I want a replacement order initiated."

[ ITERATION 1: REASONING ]
Thought: The user has a hardware issue and wants a replacement if under warranty. First, I must look up warranty status.
Action:  Invoke Tool: HardwareDatabaseAPI(action="lookup_warranty", user="alice")

[ ITERATION 1: OBSERVATION ]
Observation: Laptop serial #8841-A; Purchased 8 months ago; 3-Year Enterprise Warranty ACTIVE.

[ ITERATION 2: REASONING ]
Thought: Warranty is confirmed active. Now I should check the troubleshooting knowledge base to see if flickering is fixable via driver update before ordering hardware.
Action:  Query Data Store: IT_Troubleshooting_Store(query="laptop screen flickering Model 8841-A")

[ ITERATION 2: OBSERVATION ]
Observation: KB-104: Known firmware issue resolved by Display Patch 4.2. If patch fails, hardware replacement is required.

[ ITERATION 3: REASONING ]
Thought: I should advise Alice to run Patch 4.2 first, but pre-authorize the replacement ticket if she confirms the patch did not work.
Action:  Synthesize final empathetic response with actionable instructions and ticket ID.

Deconstructing the ReAct Loop

1. Reasoning (Thought): The agent's underlying Gemini model analyzes the conversation history, user goal, and available tools to formulate an internal rationale regarding the next logical step.
2. Acting (Action): The agent generates a structured API call or data store query to gather data or alter state in an external system.
3. Observation: The external system returns a payload (JSON response, error code, or search snippet). The agent ingests this observation into its active context window.
4. Reflection & Progression: The agent evaluates whether the observation satisfies the user's objective. If not, it initiates another reasoning cycle. Once resolved, it synthesizes the final grounded user-facing response.

---

Balancing Autonomous Agency with Deterministic Governance

A critical exam topic and enterprise leadership concern is avoiding "runaway agents" or hallucinated actions. If an agent has access to financial transaction APIs or production infrastructure, purely probabilistic decision-making introduces unacceptable enterprise risk.

Agent Platform solves this by enabling hybrid orchestration:

---

Component Table: Anatomy of an Enterprise Gen AI Agent

---

Strategic Leadership Guidance: Exam Tips & Common Pitfalls

[!TIP] Exam Tip: On the Google Cloud Generative AI Leader exam, pay careful attention to the operational verb in the question:

• If the system only needs to retrieve, summarize, and cite static information from corporate files, recommend Agent Search.
• If the system must take action, invoke external software APIs, query transactional databases, or complete multi-step workflows, recommend Agent Platform (leveraging Tools, Extensions, and Playbooks).

[!CAUTION] Common Pitfall: Never assume that agents operate purely autonomously without human oversight. Enterprise architectures mandate Human-in-the-Loop (HITL) controls for high-stakes decisions (such as credit denials, medical triage, or production database mutations). An agent in Agent Platform can execute routine tasks autonomously while systematically escalating complex or high-risk requests to human operators.