CertiTipsCRITERY FUNCTION
EspañolEnglishPortuguêsItaliano

Agents in OCI Enterprise AI

It moves from prototype to run: execution, sessions, permissions, observability and deployment.

A prototype works while someone keeps his terminal open. A corporate service also needs an identity, state, boundaries, deployment and observation. OCI Enterprise AI offers capabilities to bring agents to that environment and allows to distinguish consumption from a API from housing from complete application.

Key concepts

The Life cycle includes designing, building, testing, deploying, monitoring and improving. The run holds the execution: coordinates steps, tools, memory and faults. The framework helps to express logic but doesn't solve all the operational work itself.

In this module, Enterprise AI Agents referred to capacity to OCI Generative AIincluding OCI Responses API. It should not be automatically confused with a separate service called OCI Generative AI Agents. Check what product and documentation you are using before following a demonstration.

Module Objectives

Building platform and blocks

The platform connects access to models, development and governance. The OCI Responses API supports a OpenAI-supported interface for requests models and tools. Using SDK OpenAI with endpoint OCI does not mean that the demand will be run in OpenAI's infrastructure: the destination, authentication, project and model determine the service used.

To select a model, separate three tasks: chat generates responses. embeddings they turn content into vectors to recover candidates by similarity. and rerank Reorder them according to their relevance to a consultation. At a RAG, these candidates give context to the chat model. Rerank is an optional stage that you have to evaluate, not an agent's requirement. A lower temperature, if the model allows to configure it, reduces variation. It doesn't guarantee that the answer's true. Test the skills and regions in the OCI model catalogue.

The Mode on-demand offers shared and consumption management access as per use. The Dedicated AI Clusters Reserve compute's capacity for your models without sharing with other tenancy and billing by capacity and time. Choose by load, isolation and supported models. review quotas and life cycle without assuming guaranteed latency or indefinitely available. They are decisions about serving models, other than housing your application. Query dedicated clusters y cost modalities.

The Government combines network, identity and conduct: private endpoints delimit network access ZPR may add identity-based communication controls; IAM defines permissions over resources and set-up rangers review contents, injection attempts or personal data. They do not replace an action authorisation or eliminate all risks. The OCI's government guide puts these layers and Reg reference details his configuration and boundaries.

The studied blocks include File Search for documentary recovery; Code Interpreter to run code in an independently administered environment; Function Calling for functions that run your application; and calls a MCP for external tools. The File, Vector Stores and Container APIs support these scenarios. A temporary container from a tool isn't the same as deploying your entire corporate application. Review support by model and region OCI Responses API reference.

A project organize related resources. The memory of conversation keeps an interplay context. Long term memory allows information to be retrieved between interactions as set up. and compaction Condensed context. To sum up can lose details: check that important decisions and restrictions survive. Define retention and access before storing personal data. These capabilities are documented at Enterprise AI Agents at OCI Generative AI.

First verified steps

This section complete the lesson that was empty at the points provided, using the Enterprise AI Agents's official QuickStart, visited 28 September 2026.

  1. Select a practice comparison and check permissions IAM with the manager.
  2. It creates a project. register your OCID and check and memory retention before activates them.
  3. Configure authentication: A OCI Generative AI specific test key, or IAM authentication as an environment. A key alone does not replace access policy.
  4. Install the supported client indicated in the documentation and configure your region's endpoint.
  5. Choose a model available there and suitable for your ability to prove.
  6. Send a simple text request and confirm that you get an answer before adding tools.

The comprehensive permissions of the QuickStart are for a controlled learning environment. an actual application requires permissions adjusted to your operations. Keep credentials out of the repository and projected contents during a talk.

Step-by-step example

Our assistant should report if an order is ready and estimate the total cost of your items.

  1. The application sets up client, region, project and model. First make a text request to isolate potential login errors.
  2. He describes consultar_pedido y calcular_total as tools with narrow arguments.
  3. The API returns a request of consultation. The code of the valid user and identifier application and runs your test backend.
  4. The result is returned and associated with call. If you have to compute, the model requests the second tool and the code runs it.
  5. The application delivers the response at end of the loop, or returns a controlled error by reaching its limits.

It notes that, for its own functions, implementation continues to conduct operations. Using a managed service for models doesn't automatically convert all code into remote execution. preserves a minimum trace with identifiers, times and states; Do not publish actual orders to prove that it works.

A complete solution: business operation

The run coordinates sessions, state and execution. Models and tools are connected capabilities, and they aren't runtime itself. The observability allows to inspect the path of an request while the permissions have to be checked at each resource and operation, not only at initial access.

How to read the image: Following an request from the application, identify where your state remains and what components are involved if you have to check data. The view is conceptual: agent symbols do not represent a number of requested replicas, nor does the drawing require an OCI topology, an automatic climbing or a service availability.

Retiring and climbing

On the first track, the application lives with your laptop, a VM, a service or a cluster and consumes OCI Responses API. Your team's still operating that deployment. In the second, you prepare a compatible and public and packaging application by using managed OCI Generative AI applications accommodation.

The reference path is to develop and test, pack a image, publish it in the corresponding registry, create the application and configure an deployment with endpoint. Define network, authentication, storage and capacity as required. The Enterprise AI documentation describes the applications housed. Do not confuse replicate the server with making the models or APIs that it consumes unlimited: its quotas and latences continue to matter.

Before to update a published agent, verifies its instructions, tools and settings. Assesses the new version with normal cases, errors and boundaries. then unfolds monitoring quality, latency and malfunctions. Keep the previous version so that you can get back if regressions appear. A better response in a single example does not show an overall improvement.

Common errors

Election exercise

The computer wants to preserve its application into an existing backend and only replace the inference provider with OCI. Do you necessarily need to deploy it as an Enterprise AI hosted application?

Solution

No. You can preserve backend and consume OCI Responses API by adjusting authentication, endpoint, project, model and supported capabilities. The run accommodation is a second option when you also have to move the operation from the application. In both cases you have to test tools, state, permissions and bug management.

Sources and review

Use Quick Start to practice and OCI Responses API To check support. Before to move forward explain where every feature of your example runs and who owner is to scale, authorize and record its result.