How (Human) Developers Should Upskill in the AI Era
AI code assistants are being treated as digital interns, AI employees are becoming more common, and realistic avatars of CEOs are popping up on earnings calls. Where does this leave human developers in an increasingly AI-centric workforce?
NVIDIA CEO Jensen Huang has backed off his earlier rhetoric of AI replacing coders. AI is making everyone a coder and people don’t need to know C or C++, he said last month at a conference.
“It’s unquestionable: you’re not going to lose your job to an AI, but you are going to lose your job to somebody who uses AI,” Jensen said.
Coders need to adapt — and quickly — to the emerging job definition in a multi-agent world, which will value technical depth, business acumen and systems-thinking.
“It becomes very important for developers to understand the decisions being made and what variability might happen.”
– Craig LeClair, Forrester Research
Developers will spend less time on keyboards banging out raw deterministic code and more time crafting agent systems, orchestrating workflows, and writing effective instructions for AI models.
“I don’t think the same skills that make you a good Java programmer or C++ programmer are going to be the same skills that are going to make you a good agent builder,” said Jayesh Govindarajan, executive vice president of Salesforce AI.
Developers will spend more time plugging AI agents into operations that can make decisions autonomously. Developers need to think big and understand a business, along with its processes and functions.
“You can’t always predict the outcome when AI is making decisions. It becomes very important for developers to understand the decisions being made and what variability might happen,” said Craig LeClair, vice president and principal analyst at Forrester Research.
The New Stack of AI
There’s a redefinition of full-stack within the AI agent development model; and it’s based around solving business problems.
Process knowledge plays an important role in the development stack and coders can build value by contributing to decision making, LeClair said.
AI helps backend developers move up the stack to business logic, orchestration and frontend design. For example, ChatGPT works as a Figma tool with the generative components that allow coders to play around with interfaces, Salesforce’s Govindarajan said.
By the same token, frontend developers and designers can use AI to move further down the stack with basic backend integration, working with APIs and data connections. Protocols such as Model Context Protocol (MCP) and Agent2Agent (A2A) are becoming necessary in multi-agent systems, said Bob Parker, senior vice president for enterprise application research at IDC.
“It’s kind of like they need each other for the agents to work together,” Parker said.
Lightning-Fast Iteration
Agents are gutting the traditional software-delivery lifecycle, Forrester’s LeClair said.
“Technology is racing way ahead of the discipline we need on how to design these processes,” LeClair said.
Developers can cook faster with AI tools, and Salesforce’s developers can produce working prototypes linking the frontend and backend.
“The iteration loop is incredibly fast because we can give you something in 15 minutes,” Govindarajan said. “It used to be some janky command-line demo that an engineer would show — I love those — but it’s so much more complete now.”
“You start with the core that you’re strongest in, and then use ChatGPT, Claude and others…”
– Jayesh Govindarajan, Salesforce AI
Full-stack programming begins with a strong technical base, which could be in backend, frontend or data science. AI tools help fill technical gaps up and down the stack.
“You start with the core that you’re strongest in, and then use ChatGPT, Claude and others — you use a whole family of tools to become more end-to-end in being able to build systems that have all of it,” Govindarajan said.
A Third Pillar
Salesforce’s Govindarajan added a third pillar to the AI development stack: data science.
“We build a lot of models, we clean a lot of data, we tune them, we bring in optimizations. There’s a science aspect to it as well, which is less automated, but still being able to pull all of those three things together is the redefinition I think of full-stack,” Govindarajan said.
Learning enough science goes a long way in evaluating non-deterministic AI systems, which can easily go off track. These systems don’t offer the predictability of conventional systems.
“You can’t just say ‘hey, you gave me the wrong answer.’ You need to be able to detect that. That’s where evaluation comes in,” Govindarajan said.
Systems Approach
A strong software engineering foundation remains a cornerstone to building an efficient AI system, said Autumn Moulder, vice president of engineering at Cohere.
The company recently introduced a version of its large language model that can be self-hosted. Certain skills help build efficient AI systems for constrained computing offered by in-house servers.
“You have to have engineers: all the way from how you pre-train [and] post-train the model, into how you are building the APIs…”
– Autumn Moulder, Cohere
“You have to have engineers: all the way from how you pre-train [and] post-train the model, into how you are building the APIs, and the serving framework that calls that. And then the application itself — how is it leveraging the model?” Moulder said.
All of those things have to be tightly integrated into one efficient unit that can run in a private environment.
“Those are just all very much software engineering skills that will matter,” Moulder said.
Google provides an API stack to Gemini AI for managed services, so that users don’t have to worry about the underlying stack.
Business Processes and Domain Expertise
Domain knowledge in specific areas will help developers stand out, Moulder said.
“You have to understand the business vertical and how agents plug into the workforce,” Moulder said. “You need people who understand the business process and can say, this is what the model is capable of.”
There are about 200 startups developing low-code tools for developers to quickly create autonomous AI agents, Forrester’s LeClair said.
“Executive” agents arriving in the next few years will automate some decision-making, LeClair added.
As these sophisticated agents make their mark, developers will start stringing agentic tasks together into workflows, which then turn into processes.
“A systems thinking mentality is extremely important to understand the processes and how this fits into the big picture.”
– Stephanie Walter, Hyperframe Research
AI hallucinates, and developers will have to know when to bring humans in the loop.
Developers will also fix technical obstacles facing AI agent implementations in organizations — such as explainability, data security, guardrails, monitoring, ethics and bias.
“You’re going to have an assortment of models… you’re going to have control and governance around all of these trust factors,” LeClair said.
Developers taking ownership of processes to create AI personas for functions such as sales or HR will be highly valued, said Stephanie Walter, analyst in residence for the AI tech stack at Hyperframe Research.
“A systems thinking mentality is extremely important to understand the processes and how this fits into the big picture. That’s not necessarily an AI problem — AI magnifies it,” Walter said.
