Skip to content
back to the archive page
#AI

[1/2] AI in the SDLC: From Assistants to Agents (Category AI)

#AI #Software #Engineering #Productivity #DevEx #Management #RnD #Leadership #Economy

Today I gave this talk at Surf’s AI Boost conference. Here I will outline its core ideas and share materials for further reading. It builds on my earlier talk, Integrating AI into Development Processes at a Large Company, which set out a framework for adoption. This time I focused on agents.

I began with GitHub Copilot and Cursor, best known as engineering assistants. Those tools have already gained acceptance. Then Claude Code and OpenAI Codex arrived as agent systems for engineers; I have experimented with OpenAI Codex and liked it. This could change things substantially: if agent systems work well, we can delegate some jobs previously done by people — ||although these systems often still do not work well enough||. Anthropic and Google introduced MCP and A2A respectively to support these systems. I discussed those protocols earlier; the infrastructure is gradually taking shape.

There are economic drivers too:

  • AI investment: billions of dollars have gone into AI coding tools since 2023, and companies expect returns.
  • Pressure for efficiency: companies want faster development and lower costs.
  • AI providers promise the world: executives at customer companies expect a return on investment.

Some systems already produce impressive results, such as Google’s AlphaEvolve — see my analysis:

  • Data centre efficiency: it developed a heuristic for the Borg orchestrator that continuously recovers an average of 0.7% of Google’s global computing resources.
  • Chip optimisation: it proposed a Verilog rewrite that removed redundant bits from matrix-multiplication arithmetic, incorporated into a new TPU version.
  • Faster LLM training: it sped up a key Gemini component by 23%, reducing overall Gemini training time by 1%.

Researchers are considering the implications:

  • Stanford’s Future of Work with AI Agents asks which work can already be delegated to agents, what people want to delegate and how labour markets could change; see my analysis.
  • Google’s Virtual Agent Economies considers an economy in which agents perform and coordinate work autonomously, how it might affect the existing economy and how to manage the transition; see my analysis.
  • Towards AI-Native Software Engineering (SE 3.0) describes a third phase of software engineering: after conventional coding and coding with assistants, agents enable intent-first development. My review is coming, but you can read the original paper now.

For an example of agent mode in practice at T-Bank, see the demos in my talk and my colleague Stas Moiseev’s published Big Tech Night talk, which I discussed here. Agent mode alone will not deliver AI’s benefits: we need to focus on the actual jobs engineers perform. I will cover that, and measurement of the results, in the next part.

#Software #Engineering #Productivity #DevEx #AI #Management #RnD #Leadership #Economy