Prepare and test your skills
Prepare and test your skills
Business Transformation through Autonomous Workflows is the shift from AI that only answers questions to AI agents that carry out complete business processes with little human help. An autonomous workflow is a process where an agent performs multiple steps and makes routine decisions to reach a goal. This change affects who does the work, how fast work gets done, and where people spend their time.
Static retrieval is the behavior of generative AI that finds or creates information but stops at an answer. An AI agent adds goal-directed execution: the agent interprets a request, gathers context, chooses actions, and uses tools or business systems to finish the work. Data flows from systems to the agent for context, and actions flow from the agent to those systems to complete tasks. The business result is that automation can move beyond chat and search into real process execution.
How agents reduce human intervention is by taking over repetitive steps, data handoffs, and routine decisions inside multi-step workflows. Instead of a person moving information between systems or choosing every next step, the agent follows rules and uses context to keep the process moving. For complex decisions, the agent evaluates options and acts within defined limits. People then handle exceptions, approvals, and oversight rather than every step.
A typical autonomous workflow follows an ordered pattern:
A left-to-right process flow showing trigger, gather context, plan and decide, execute actions, and monitor and escalate, with the middle three steps owned by the AI agent.
The agent owns the middle steps, while the business keeps control over boundaries and exceptions.
Business implications of end-to-end automation are faster processes, lower manual effort, and more consistent operations. Agents can work continuously without waiting for manual handoffs, so cycle times shrink. Because the same rules and context are applied to similar cases, decisions become more consistent and operations can scale without adding staff at the same rate.
These benefits depend on governance. The business must set clear limits on what the agent may do, how it accesses data, and when it must ask a person. Without monitoring and escalation paths, automation can spread errors or break policy.
Exam tip: Static generative AI retrieves or generates information; an AI agent goes further by planning and executing actions to complete a multi-step task.
Foundations of Agentic Reasoning and Autonomy explain how an AI agent works toward a goal on its own instead of answering one question at a time. An AI agent is a system built on a large language model that can interpret a goal, make a plan, and carry out multiple steps while tracking context. Decisions flow from the agent to follow-on actions until the goal is complete, rather than stopping after a single answer.
The large language model (LLM) is the reasoning core of an AI agent that understands language, interprets intent, and decides what to do next. The LLM receives input from the user goal and from prior results, and it produces reasoning and next-step decisions that flow from the model to the planning function. Because the LLM can interpret open-ended instructions and weigh options, the agent can handle complex, goal-oriented work instead of only giving fixed answers. Without the LLM, the agent could not understand context or adapt when conditions change.
Planning capabilities are the processes by which an AI agent breaks a goal into ordered steps and adjusts that order as new information arrives. The planning function depends on the LLM for reasoning, and execution depends on the plan to know what to do first, what to do next, and when to stop. First the agent interprets the goal, then it creates a sequence of sub-tasks, and only after each result returns does it decide whether to continue, revise the plan, or finish. This loop of plan, act, and review gives the agent autonomy across multi-step work.
Memory is the component that lets an AI agent maintain context across a complex, goal-oriented interaction. Memory stores information that flows from earlier user messages, prior steps, and intermediate results to later decisions, so the agent does not treat each turn as a new conversation. Short-term memory holds the current conversation and active task state, while longer-term memory can retain user preferences, past outcomes, and key facts across sessions. Because later reasoning depends on this stored context, the agent can stay consistent, avoid repeating work, and build toward the final goal.
Tool Integration and Function Calling Mechanisms are how an AI agent reaches beyond its own model to do real work with enterprise data. The agent queries databases, calls APIs, and pulls verified business context so its answers are grounded instead of guessed. In Google Cloud, this happens through AI functions that the agent invokes inside BigQuery and through Knowledge Catalog (opens in a new tab) that supplies the semantic context for choosing and using those functions correctly.
How agents invoke functions in BigQuery is by generating SQL from a user's question. When a user asks a high-level question in conversation, in a verified query, or by using keywords, the agent translates that intent into SQL that calls the right AI or ML function. A verified query, such as one a data scientist prepares for a recurring report, can also embed supported functions to describe defaults and automate that report. Running generative AI queries requires the appropriate permissions.
The functions an agent can call include:
The choice between search functions depends on table configuration. The agent uses AI.SEARCH only on tables that have autonomous embedding generation enabled. On any other table, it falls back to AI.SIMILARITY, which generates embeddings in real time.
Exam tip: AI.SEARCH requires a table with autonomous embedding generation enabled; without it, the agent must use AI.SIMILARITY, which generates embeddings in real time.
Example prompts that trigger functions show how the shape of a one-shot request maps to a specific function and dataset. The agent selects the function based on what the user asks for, such as forecasting, anomaly detection, or semantic search.
| Use case | Sample prompt | Dataset |
|---|---|---|
| Forecasting | "Predict the number of trips for the next month." | bigquery-public-data.san_francisco_bikeshare.bikeshare_trips |
| Anomaly detection | "Find outliers in trips per day for 2018 using 2017 as a baseline." | bigquery-public-data.san_francisco_bikeshare.bikeshare_trips |
| Key drivers | "Identify the key drivers for changes in trip volume between 2017 and 2018." | bigquery-public-data.austin_bikeshare.bikeshare_trips |
| LLM text generation | "For each article in the 'sports' category, summarize the body column in 1-2 sentences." | bigquery-public-data.bbc_news.fulltext |
| Semantic search | "Find the top 5 reviews that most closely match 'tension-building psychological thriller'" | bigquery-public-data.imdb.reviews |
Each row connects a business question to the data the agent must query. The agent depends on that mapping to generate the correct SQL and function call.
Knowledge Catalog as the context layer is the semantic orchestration plane that gives agents the business meaning behind enterprise data. Knowledge Catalog connects to operational databases, warehouses such as BigQuery and Spanner, unstructured stores such as Cloud Storage, and runtime metastores such as Lighthouse. It then exposes grounded schemas to AI models and applications so the agent can query correctly.
Knowledge Catalog distinguishes between three kinds of data that agents must handle:
txn_qty to purchased quantity.For unstructured sources, discovery scans extract entity relationships into a graph profile of AI-extracted nodes and edges. For structured sources, Knowledge Catalog crawls metadata and records columns and constraints.
Context orchestration and delivery to agents is the ordered workflow that moves verified business context from source systems into the agent's prompt. The agent depends on Knowledge Catalog for context, on delivery interfaces for transport, and on orchestration frameworks to place that context where the model can use it.
This flow turns a static model into an agent that can act. Context flows from source systems to Knowledge Catalog, from Knowledge Catalog to the agent through MCP or LookupContext, and from the orchestration framework into the prompt before the model chooses a function.
Integration with related Google Cloud services shows how Knowledge Catalog works with storage, warehousing, and modeling services to form the data foundation for agents.
| Service | Primary role | Relationship to Knowledge Catalog |
|---|---|---|
| Knowledge Catalog | Governance and agentic context | Serves as the unified semantic context graph, runs data quality scans, and exposes grounded schemas to AI models and applications. |
| BigQuery | Enterprise data warehousing | Stores, queries, and processes datasets; Knowledge Catalog crawls, indexes, and enriches these tables with business classification aspects. |
| Vertex AI | AI model development | Builds, deploys, and hosts foundation models; custom agents retrieve metadata context from Knowledge Catalog to ground logical reasoning. |
| Cloud Storage | Unstructured file storage | Stores raw files; Knowledge Catalog runs unstructured discovery scans to parse PDFs and images and build relationship maps. |
Knowledge Catalog also cross-links with the Gemini Enterprise Agent Platform to feed tools and context into that platform's Agent Registry. It supports agent profiles such as data discovery agents, which analyze long-form intent to retrieve relevant physical assets rather than relying on keyword matching.
Static retrieval finds or creates information but stops at an answer, while an AI agent adds goal-directed execution by interpreting a request, gathering context, choosing actions, and using tools or business systems to finish the work. Data flows from systems to the agent for context, and actions flow from the agent to those systems to complete tasks, which moves automation beyond chat and search into real process execution.
An AI agent reduces human intervention by taking over repetitive steps, data handoffs, and routine decisions inside multi-step workflows and by evaluating options and acting within defined limits. The workflow follows a trigger that starts the process, context gathering, planning and deciding, execution through tools, APIs, or business applications, and monitoring with escalation of uncertain or out-of-limits cases to a human, so people handle exceptions, approvals, and oversight rather than every step.
Use AI.SEARCH only on tables that have autonomous embedding generation enabled, and use AI.SIMILARITY on any other table. AI.SIMILARITY finds semantically similar content and requires real-time embedding generation, while AI.SEARCH performs semantic search only on tables with autonomous embedding generation enabled.
Knowledge Catalog is the semantic orchestration plane that gives AI agents the business meaning behind enterprise data by connecting to operational databases, warehouses such as BigQuery and Spanner, unstructured stores such as Cloud Storage, and runtime metastores such as Lighthouse. It links metadata to semantics, structured data, and unstructured graph profiles to assemble an active context graph, which is retrieved through MCP or direct LookupContext API endpoints and injected into large language model (LLM) prompts by orchestration frameworks such as the Agent Development Kit (ADK) or LangChain to ground reasoning without hallucinations.
An enterprise is developing an autonomous AI agent to resolve multi-step customer inquiries without continuous human intervention. The engineering team needs to design an architecture that allows the system to comprehend goals, break down complex requests into sequential steps, and retain conversational state across long interactions.
Which combination represents the core components required for this autonomous agent architecture?
Last updated