GradientAI Platform, Now Generally Available

This is a great question, and one that the industry is still figuring out how to answer. Agents sometimes seem synonymous with LLMs and conversational interfaces like ChatGPT. While others take a more strict definition, where an application must have “agency” in order to qualify as agent—i.e. the ability to act independently based on input from a non-human source. At DigitalOcean, we think of agents as more complicated than the LLM itself. It’s an application that is powered by an LLM but also incorporates stateful components like conversational memory and training on one or more data sources (a vector database, or “Knowledge Base”). It may have a human-like role or personality, and it may have a conversational interface, but it may not. We think of agency as a spectrum of autonomy ranging from simple rule-based responses to complex, goal-oriented, and self-improving behavior.

GradientAI Platform simplifies the process of integrating the components of AI Agents and LLM applications, such as prompt management, prompt evaluations, data sources, 3rd-party tools, and conversational memory. Whether you just need simple LLM API calls or you’re building complex, stateful applications, GradientAI Platform makes the process significantly faster than working directly with the individual components. It gives you deep insight into agent performance. And it also simplifies account management and billing, with all costs coming to you on a single invoice, even if you switch LLM models. Plus, you don’t have to host anything yourself or bring your own API keys, like with similar frameworks.

Visibility into agent performance is one of the key value propositions of GradientAI Platform. Think of our Agent Evaluations feature as the test suite in a traditional code project. It allows you to create a set of prompts (like unit tests), and run your agent against them after any change you make to its instructions, knowledge base, functions, routing, or LLM model. Our Traceability feature exposes the reasoning chain behind an output, including user inputs, model outputs, and function call results. If you don’t like the response to one of your test prompts, or you hit an error along the way, Traceability can help you identify the source of the problem. Assuming the agent has a user interface, you can also view feedback data from users who rate the agent’s responses. Conversation Logs (coming soon) allow you to dig into real user conversations to understand how agents are responding. And Citations help you locate which training data informed the agent’s responses.

At DigitalOcean, your data is never used to train our AI or systems, and it’s never sold or shared. If you select open-source models like Mistral, Llama, or DeepSeek, your data stays entirely within DigitalOcean’s infrastructure, and we store only what we need to power your agent or to provide you with troubleshooting features like Conversation Logs.

You have several options for interacting with the agent at every phase of the Agent Development Lifecycle. Within the DigitalOcean console, it takes about two minutes to spin up an agent with an initial set of instructions. You can immediately start prompting the agent from the Agent Playground and make adjustments as necessary. For ongoing testing, Agent Evaluations provides real responses to a set of prompts or benchmarks that you define, and you can choose how often to run your evals. For integration with broader applications, your agent comes with its own API endpoint and endpoint access key. If you want the agent freely accessible to outside applications without your access key, you can set the endpoint status to “public”. Check out this example of integrating a GradientAI Platform agent with Telegram.