AI in SDLC: from assistants to agents
Copilots, MCP/A2A, AlphaEvolve, SE 3.0 and large fintech cases
Slide contents
1. AI in SDLC: from assistants to agents
Copilots, MCP/A2A, AlphaEvolve, SE 3.0 and large fintech cases
2. Alexander Polomodov
Technical Director & Fellow, large fintech
Architecture and engineering practices in fintech
AI in SDLC across thousands of engineers
Podcast «Code of Leadership», @book_cube
3. From tools to measurable outcomes
Why agents emerged, how to build them, and how to measure their effect
Why agents became possible
How engineering work changes
How agents enter the SDLC
How impact differs from usage
4. 01. Previously on this show
How we integrate AI into SDLC — big tech cases
5. Agents continue the tooling evolution
The previous talk covered AI integration; this one follows the autonomous loop
Assistants already live in tools
Autonomy is the next step
Agents need a shared platform
Outcomes still need proof
6. 02. Assistants → agents
From autocomplete to autonomous task execution
7. Assistants became everyday tooling
Copilot and Cursor opened the path; Claude Code and Codex pushed agent mode
Copilot: agent mode
Cursor: AI-centered IDE
Claude Code: terminal agent
Codex: long-running tasks
8. An agent owns the task loop
The human states the outcome and decides how deeply to inspect the result
Plans its own steps
Calls external tools
Acts without step-by-step commands
Returns a result to the human
9. 03. Protocols: MCP and A2A
Standards for agents and integrations
10. MCP connects tools; A2A connects agents
The protocols solve adjacent problems and can work together
11. 04. Research and economics
Why agents are real and here to stay
12. Three forces push the industry toward agents
Investment demands returns, business demands efficiency, and the market expects promised outcomes
Billions invested since 2023
Pressure on company efficiency
Promises from AI providers
13. AlphaEvolve improves programs through an evaluator
An LLM proposes a change, a metric tests it, and the database retains better versions
14. A measurable evaluator turns search into impact
Google applied one loop to infrastructure, hardware, and model training
Borg scheduling: +0.7%
TPU circuit optimization
Faster matrix computation
Gemini training: +1%
15. People do not want every feasible automation
Compare worker desire with the actual capability of current agents
Green light — Wanted by workers, feasible today
R&D — Wanted by workers, not feasible yet
Red light — Feasible today, unwanted by workers
16. Agent economies need deliberate rules
The authors prefer designed sandboxes over an emergent, permeable market
Low risk
Permeable sandbox
Rules designed up front
High risk
Bounded sandbox
Rules designed up front
17. 05. SE 3.0 — AI-Native Engineering
A new paradigm of software development
18. SE 3.0 moves the interface from code to intent
An AI-native process connects human requirements with machine implementation
SE 1.0: code-first + tools
SE 2.0: code-first + AI
SE 3.0: intent-first + dialogue
Humans define requirements
19. Start with the developer goal, then choose tools
The developer journey links SDLC stages, durable goals, and platform teams
Gather information
Plan and gain approvals
Develop, test, and commit
Release and operate
20. 06. Large fintech case
Spirit and agent mode in production
21. Agent-first starts with the whole platform
Spirit spans work intake and knowledge through infrastructure and production readiness
Process, search, and documentation
VCS, IDE, CI, and builds
Delivery, cloud, and components
Observability and incident management
22. Agents must cover the whole SDLC
The engineer sets direction, designs the system, and reviews the output
23. 60% of time is not code writing
Web and mobile team analysis: an AI program cannot stop at autocomplete
Creation
Code — ~40%
Tests — ~15%
Coordination
MR / code review — ~15%
Jira — ~12%
Wiki — ~12%
24. Agents move beyond the IDE
Teams choose painful jobs across the lifecycle
Code: complete a task end to end
Data: request → code → chart
QA: description → test case
Each profession chooses its jobs
25. 07. Agents in production
T-Cover Agent, AI code review, security
26. T-Cover proposes only validated tests
Change → coverage need → candidates → validation → MR for a developer decision
27. Lower baseline coverage yields larger gains
Added coverage declines as the existing test base grows
Up to 66%
0–20% → +40 pp
21–47% → +33 pp
47–66% → +27 pp
Above 66%
66–73% → +23 pp
73–89% → +12 pp
28. AI review starts with context
GitLab, docs, Jira, and instructions feed classification and comment ranking
29. A security agent sees code, not every control
SafeLiner finds a vulnerability and proposes a fix, but may miss infrastructure mitigations
Scans the repository
Explains the finding
Offers a one-click patch
May miss infrastructure protection
30. 08. Measuring impact
Usage is only the first level of evidence
31. Utilization starts; impact still needs proof
Measure usage first, then impact, then cost and return
32. Usage grew; impact still needs proof
Internal data clearly establishes adoption and retention
Created artifacts
IDE H1: 10%
IDE H2: 25%
Heli: 5% → 10%
Retention · 4 weeks
IDE: suggest 79%, chat 55%
Heli: suggest 70%, chat 46%
33. The Russian market needs a comparable baseline
The study connects practices, engineer perception, and measurable outcomes
Practices across Russian companies
Quality and efficiency
How engineers perceive AI
Comparison with global trends
ai4sdlc.tbank.ru
34. Agentic development is a system
Five conclusions from research and production cases
Protocols connect the ecosystem
Evaluators turn search into outcomes
Platforms span the entire SDLC
Engineers state intent and review
Utilization is not impact
35. Sources behind the argument
Research on discovery, work, agent economies, SE 3.0, and measurement
AlphaEvolve · Google DeepMind
Future of Work · Stanford
Virtual Agent Economies · Google
SE 3.0 · Measuring AI · DX
36. Thank you!
polomodov.tech
All slides and links are in the Telegram channel
Alexander Polomodov, Technical Director & Fellow, large fintech
@book_cube