Agentic AI is emerging as a transformative force that promises to redefine software development and the entire structure of knowledge work. Dev Interrupted recently hosted Amir Behbehani, a mathematician, AI/ML expert, and chief AI engineer at Memra. We explored how these intelligent agents are changing how we approach complex tasks and what this means for the future of work.
How agentic AI transforms workflows beyond automation
Agentic AI represents a significant evolution beyond traditional automation tools. Unlike previous technologies such as Robotic Process Automation (RPA), which primarily execute predefined tasks with specific data inputs, agents operate with higher reasoning and adaptability.
These agents integrate several critical components: access to large language models (LLMs), long-term memory (similar to vector and graph databases), short-term memory for adaptability, and specialized training for specific tasks. Their power comes from their ability to reason through complex imperatives, which Behbehani describes as a "Socratic reasoning" process: asking questions and breaking tasks into manageable components.
Most importantly, these organizations increasingly view AI agents as digital employees who can perform specific organizational roles. As Behbehani notes, "These agents are effectively digital employees, and those digital employees paired with a human employee, say an individual contributor, suddenly that individual contributor has multiplicative returns to scale."
This "multiplicative" effect comes from shifting humans from being in the middle of workflows (maintaining context and performing individual tasks) to being at the top of workflows (providing high-level direction and oversight). The agent handles the complex reasoning and execution processes, allowing humans to focus on more strategic work.
How AI redefines roles from coders to architects
The integration of AI into software engineering is already happening at a remarkable pace. Research from LinearB found that over 13% of pull requests are now created by bots, and this percentage is expected to grow significantly as agentic AI continues to develop.
The evolution of software engineering is moving from an execution mindset to a management and architectural one. Engineers increasingly become managers of AI systems rather than executors of specific coding tasks. Behbehani describes this as shifting from a "carpenter" mindset to an "architect" or "industrial engineering" mindset:
"If you're building agentic frameworks and then deploying those agents to effectively write code... you're building the systems that effectively build the constituent systems that build the application."
This approach represents a fundamental shift in how engineers approach their work. Rather than focusing solely on writing code, engineers must design systems that can generate code and applications. They'll need to think about the inputs they give to agents, how to manage information flows, and how to architect complex systems that AI can build.
The economies of scope of agentic AI
As AI agents become more capable, they're beginning to automate workflows by breaking them down into constituent tasks, creating opportunities and challenges for human workers across industries."Agents offer economies of scope. The way to think about an agent is you're designing a tooling line that can effectively manufacture various products," Behbehani explains.
This capability means that agents can climb what Behbehani calls the "career ladder," starting with basic tasks like data entry and booking meetings and progressing to more complex functions like writing applications or managing entire workflows. Though he notes we're still "far from the agents necessarily solving for product-market fit" or handling the most complex creative tasks, the trajectory is clear.
The implications for labor markets are profound. Traditional role-based work structures may become too rigid when agents can fluidly handle multiple tasks across domains. This could lead to a fundamental rethinking of how work is organized, with humans focusing more on relational structures and ownership of work products rather than execution of specific tasks.
Emerging AI frameworks and vertical integration
The development of AI is following a different pattern from previous technological waves. Rather than being built from the top down (starting with applications and then adding AI capabilities), AI is developing from the bottom up:
"The foundational layer comes about first, then the rag layer, then the short-term memory layer, and then these agentic frameworks. The application layer is still around the corner," Behbehani notes.
This bottom-up development pattern creates interesting dynamics for both startups and established companies. Behbehani strongly advocates vertical integration in AI development, where companies control multiple layers of the AI stack to capture crucial feedback loops that improve their systems.
For startups, this means focusing on rapid innovation while solving the persistent distribution challenge. For established companies, it means finding ways to insource AI capabilities and vertically integrate them into existing operations before more agile AI-first companies can disrupt them entirely.
Navigating the AI future beyond code copilots
For individual engineers and engineering managers wondering how to navigate this changing landscape, Behbehani offers two approaches:
- Leverage AI tools for marginal productivity improvements (20-30%)
- Focus on designing the agentic AI systems that will take development beyond code copilots by "building the systems that build systems."
However, the most crucial advice is to engage directly with these technologies. As agentic AI continues to develop, we'll likely see the emergence of what Behbehani calls "synthetic marketplaces." These platforms describe work and provide agents to solve the tasks, similar to how Waymo provides ride services without human drivers. This represents another way AI may transform traditional labor markets in the coming years.
The future of software engineering with AI appears to be one where humans remain in the loop, contributing meaningfully to codebases while benefiting from AI assistance. At the same time, a parallel stream of development will focus on building frameworks that further automate development processes. The balance between these approaches - what Behbehani calls organizations with "co-pilots versus autopilots" - will likely define the next era of software engineering.
Listen to Amir’s full Dev Interrupted episode here:
