[1/2] AI in SDLC: the way from assistants to agents (AI column)
Today I made such a presentation on AI Boost Conference from Surf. In this post, I will talk thesis about the very essence of the speech, and also share all the materials for additional study. And to begin with, this talk continues my previous one.Integrate AI into development processes in a large company", which set the frame and well showed how and where to go when implementing AI. But this time I focused on agents. And I started by thinking about GitHub Copilot and Cursor, which are more commonly known as engineering assistants. And the adoption of these tools has already happened, but then came Claude Code and OpenAI Codex, which declared themselves as agent systems for engineers. (I experimented with OpenAI Codex and I liked it.). And it seems that this is a gamechanger, because if agent systems work well, then they can delegate some of the jobs to be done scripts that people used to do. (||These systems often don’t work well enough.||). To improve the performance of these systems, Anthropic and Google came up with the MCP and A2A protocols, which I have already developed. toldThe infrastructure is gradually being built.
There are economic prerequisites for the transition to agents. - Investment in AI: 2023 Billions of dollars have been invested in AI programming tools, and companies expect returns - Pressure on efficiencyCompanies are looking for ways to speed up development and reduce costs AI technology providers promise Manna heavenly: top managers in consumer companies are waiting for return on investment
And even there are companies where these systems give top-level results - it's Google with their agency system. AlphaEvolve (see. my analysis)where tangible results have been achieved
- Improving the efficiency of data centersAlphaEvolve has developed a heuristic for the Borg orchestrator that continuously restores the average. 0,7Percentage of Google’s global computing resources. - Chip optimization: The system proposed to rewrite the Verilog code to remove excess bits in the arithmetic matrix multiplication scheme, which was integrated into the new version of the TPU.
- Accelerating learning of LLM modelsAlphaEvolve Accelerates Key Component of Gemini Architecture 23%, which led to a reduction in Gemini training time 1%
Against this background, respected scientists consider questions of the future, for example.
- "Future of Work with AI Agents" from Stanford, where the kids asked what can already be agencyd, and also what people want to pass on to machines, and also how the labor market will change as a result. (centimeter my analysis)
- "Virtual Agent Economies" from Google, where the guys were thinking about an economy where agents can do and coordinate work autonomously." Scientists were interested in how this would affect our familiar economy and how to make sure that we could manage these changes. (centimeter my analysis)
- "Towards AI-Native Software Engineering (SE 3.0)"where the authors speculate about changing the development industry to a third version, where the first was standard code writing, the second was writing code with assistants, and the third is already agents in all their glory with intention-first development." This article is on the way, but for now you can read it. original
What an agency regime might look like in reality (We're in T Bank.) It is worth looking at the demos, which are in my report and in the already published report of Stas Moiseev, my colleague, who spoke about this and not only at Big Tech Night. (centimeter my analysis). But to get the effect of AI, simply agent mode is not enough – you need to work focusing on jobs to be done scenarios that engineers perform as part of their work. More about this, as well as the measurement of the effect will continue.
#Software #Engineering #Productivity #DevEx #AI #Management #RnD #Leadership #Economy