AI Coding Agents Will Promote You To Tech Lead by Ed Lyons

Image by Ed Lyons via Midjourney

Last week’s ODSC East Conference in Boston was all about the potential of AI agents, and a great opportunity to inhale the hype, and then hear about the successes and difficulties in using them. It was also a chance to hear from very different people about their experiences and perspectives. While all kinds of agents were discussed, I was most interested in coding agents.

It was certainly a shock to hear a vice president of SAP casually say in a keynote address that his company’s “army of software engineers are going to have to learn something new.” Soon after, it was then quite different to hear other speakers from the same podium talking about the enormous and difficult human software engineering efforts underway to enable agent-enabled applications inside organizations. 

It was illuminating and challenging to hear from someone working on building a coding agent talk about using agents for that work. His sober message was both “coding is going away” but that software engineers are still needed. 

And what a breath of fresh air for a professor to debug several other conference talks by showing serious research that tested the abilities of the latest AI tools. He then predicted that fully using AI inside businesses will take “decades.” I thought he might be escorted out of the convention center for saying that. Alas, he remained.

And yet another speaker spoke of a new world where we can generate more software than we have ever considered, such as no longer having one product for all customers, but every customer having their own customized version.

What a confusing time we live in! How should we think of AI’s ability to write software? Will it replace software engineers, as so many are predicting? If not, how should organizations view the practice of software engineering and the role of these powerful agents? How can this technology be adopted wisely?

Agent-based coding tools are tremendously powerful, and no discussion about their problems should steal too much attention from what they can do. Everyone who can use them, should use them.

But the sobering research findings from Princeton Professor Arvind Narayanan were a reminder that hallucinations are a problem and they are not going away. He also said that there is a “leaderboard” mentality among vendors of AI tools, and that the benchmarks that determine your ranking do not predict success in real world usage. He gave an example of the headline-grabbing AI software developer named “Devin” that looked great on the benchmarks, but failed in most real-world tasks.

Other speakers pointed out that having agents do lots of steps unchecked, or enabling autonomous cooperation between agents, simply multiplied the chance of mistakes along the way. I have seen this in my own agent use. Allowing my agent to do things step by step so I can check how things are going can prevent both errors and architectural problems. This is something to keep in mind for coding agent product announcements this week from Microsoft and OpenAI. The more they do on their own, the greater the need for oversight. 

My biggest take away from Dr. Narayanan’s talk was his idea that “adoption is the speed limit on innovation.” He said that despite the power of AI, adoption is an enormous issue and companies will take a long time to process AI. He gave the example of self-driving cars, which are very good at driving, yet adoption problems have prevented large-scale use for years. This means that there is time for software engineers to adjust. But since software development is adopting AI faster than other fields, it will not take years to see major changes.

Robert Brennan, CEO of All Hands AI, which makes the quite-capable OpenHands coding agent, gave a great presentation. He said that they use their own agent for their software and it generates about 80% of the code they use. If even the agent specialists cannot generate it all, then your organization is not going to do so either, and will likely generate significantly less than that. But Brennan did say, with no conference-style excitement, “coding is going away.” He also said that management of these agents is crucial, and is still real engineering. Still, there will be a lot less typing of code. Brennan said a lot of the resistance to agent-based AI is that while previous coding tools were in-line and complimented existing workflows, using agents is very different and many developers are uncomfortable with such a big change to long-standing habits.

I spoke to him after his presentation to follow up. We spoke of a few issues, including the important distinction that generating “80% of the code" is certainly not most of the work, and that code certainly did not happen all at once. As far as the role of today’s software engineers, the conclusion we agreed upon was that the job of “programmer” is going away. That what is needed instead are many “technical team leads” where a former developer is reviewing a lot of code he or she did not write, and they must be concerned about the architecture and what their customers ultimately want. They have to drop in and write code once in a while, but this is not their primary task. Talking to engineers (AI and human) about direction, system characteristics, and customer needs - and making sure the application is in alignment - is what their job is in the world of agents. 

Many experienced software developers have remained individual contributors and are not currently playing the role of tech lead, but coding agents will now be their direct reports. Like it or not, human engineers will have to talk to the agents daily, correct them, and make sure they do a good job. These human developers are now faced with a choice: either be in charge of the agents, or compete with them directly. 

I’d take the promotion while it is still being offered.

For companies wondering about how to adjust to this new world of software development, I would recommend that they see agents as the ability to add mid-level programmers who work really fast but make mistakes, and often start coding before they have all the direction they need. Your existing developers need to let go of a lot of what they were typing, let these new developers take over a lot of that coding, and get used to new workflows.

But it will be well-worth the effort. Your organization will be able to move much faster creating software with agents if your engineers can handle the change in role, learn agents’ strengths and weaknesses, and get comfortable in managing them.