AI Engineering Podcast

What Do Humans Do Now? Rethinking Software Engineering in the Age of AI Agents

1h 8min

Jonathan Ellis argues that with AI agents generating code 10x faster, traditional human code review is no longer a good use of time; the future of software engineering shifts to human judgment in design decisions, product management, and building automated quality assurance (oracle) systems. He emphasizes investing human effort in creating validation tools, skill documents, and agent harnesses, while moving from line-by-line review to higher-level oversight of agent workflows.

Summarized by Podsumo

🎧 Listen 🎙️ Ask about this episode

✨ Key Takeaways

1

Ellis advocates for eliminating human code review, as it catches only ~30% of defects and is outpaced by AI code generation speed.
2

The role of engineers is transforming to product managers—deciding what to build, not how to build it—with agents handling implementation.
3

Building automated oracles (e.g., fuzz testing, property checkers) and agent-friendly documentation (e.g., exec plans) is a high-return investment for human time.
4

Using a meta-harness (like MuleNir) to coordinate multiple AI subscriptions and subagents can unlock cost savings and parallelization.
5

The biggest tooling gap is agent-oriented collaboration: systems that automatically link tickets to prior agent sessions for regression fixes.

💬 Notable Quotes

"If you have a team of human engineers and your code reviews take the form of saying, 'I disagree with how you wrote this switch statement'… that's not a productive use of code review time."
"It's much more important to decide what to build next than how to implement it. That's the human value add."
"We should be building an oracle for our agents—something that knows the right answer for a certain class of problem—and use agents to build that oracle."