AI – Here to Help the Humans Avoid Tedium As They Build Agents

At Dreamforce in September 2024, Salesforce debuted Agentforce for Developers, the next iteration of Einstein for Developers. The suite of AI-powered tools helps developers write Salesforce-specific code, like Apex and LWC, and a new feature called Dev Assistant uses a conversational AI coding assistant to write new code, explain existing code, generate test cases, and generate documentation.

When Agentforce for Developers was introduced, one of the selling points was the idea that it would automate out tedious work – so No Jitter asked Alice Steinglass, EVP and GM, Salesforce Platform, how the company draws the line between the work that should be automated and the work that still requires human review. Steinglass further explained what these tools do for both entry-level AI-powered agent builders and seasoned developers.

Answers have been edited for clarity.

No Jitter (NJ): What is one example of how Agentforce for Developers automates tedious work, and why someone using Agentforce might want that automated instead of done by humans?

Alice Steinglass (Steinglass): Developers building applications and agents can work with Agentforce for Developers to help automate the task of generating new code or tests. Using the inline code autocomplete feature, for example, developers can receive AI-powered code suggestions directly within the editor, helping to streamline the development process. 

Developers using Agentforce can use it to help kick off the development process, working with Dev Assistant, our new conversational AI-powered code assistant, to bring code to completion faster using their data, then bringing humans into the loop to ensure the generated code is working as needed or desired. Dev Assistant can generate code for you from a single prompt, grounded in the customer's data and metadata. Ultimately, using a tool like Agentforce for Developers brings more people — such as beginner developers – into the development process using natural language, allowing businesses to build and test applications and agents more efficiently.

NJ: Agentforce for Developers has guides that can carry human users through the entire app- and agent-building process — how easy will it be for skilled users to eventually devise their own workflows or customizations within the agent-building process?

Steinglass: While the AI-assisted tooling makes it easier for new developers to onboard to the Salesforce Platform, expert developers can unlock even more productivity by seamlessly integrating Agentforce for Developers into their development workflow, providing helpful assistance for even complex coding tasks.

Additionally, Agent Builder allows customers to use natural language to create instructions and guardrails for their agents. If you can describe it, Agentforce can do it. That allows non-technical users who may be closer to the business needs to describe exactly what they want out of an agent and begin to build actions.

NJ: Dev Assistant is being positioned as a tool that can assist with tasks like writing new code. What kind of code? And on a code review, how much of the AI-generated code is immediately useful?

Steinglass: Agentforce for Developers supports Salesforce-specific code, including Apex and LWC. 

Dev Assistant leverages Retrieval Augmented Generation (RAG) techniques which enhance the relevance of responses by considering the current schema and metadata from each local project. This ensures that the tailored code suggestions from Dev Assistant are both aligned with a developer’s immediate coding needs and contextually relevant to their Salesforce org, to help make AI-generated code immediately useful. And, with Dev Assistant, you can iterate on the outputs to get the exact code you need, which wasn't previously available with the natural language-to-code functionality.

We prioritize accuracy, precision, and recall in our models and back our AI-generated code with explanations and sources whenever possible to help make as much of the code immediately useful to developers as possible. As always, we recommend that a human reviews the model output for accuracy and safety before sharing it with end users.

NJ: To follow on, Dev Assistant is being positioned as a tool that can create documentation — what kind of documentation? And how is this documentation reviewed for usefulness and accuracy?

Steinglass: With Dev Assistant, developers can generate documentation for a specific code block in the default documentation scheme for the current file’s programming language. Once code documentation is complete, human developers can quickly review the generated outputs for usefulness and accuracy before sharing them for end use. This is a big time saver for developers and allows them to focus on more critical tasks.

NJ: Finally — what was Salesforce's strategic intention for creating Agentforce for Developers? What customer problem is this addressing and how it is addressing it?

Steinglass: Customers are under pressure to build applications and agents faster to deliver ROI from their AI investments. Building those applications and agents typically falls on the developers, so its critical we create tools that empower developers to be more productive and deliver quality code faster. Enhancing developer productivity with assistive capabilities and integrating more advanced automation features is our goal with Agentforce for Developers. 

For customers who are building Agentforce custom actions or applications, or optimizing existing code, this helps them quickly and safely generate high-quality code embedded with our best practices in mind. Agentforce for Developers uses proprietary LLMs, including CodeGen2.5 and xGen-Code, to simplify complex development tasks, automate work, and guide developers through the process of building applications and agents faster.