Software development in 2030
Testable hypotheses instead of a linear forecast
Slide contents
1. Software development in 2030
Testable hypotheses instead of a linear forecast
2. Alexander Polomodov
Technical Director & Fellow, large fintech
Architecture and engineering practices in fintech
AI-native development and Platform Engineering
Code of Leadership podcast, @book_cube
3. The future resists linear extrapolation
4. 01. Signal: agents become work units
But an autonomous task is not autonomous development
5. From one agent to an agent system
6. Protocol compatibility does not create trust
7. Adoption cannot prove that delivery has improved
Tool popularity is input, not outcome
No provenance for '70% use it daily'
Generated code may increase review volume
Local speed does not guarantee system gains
We need delivery, quality and outcome metrics
8. 02. Research offers signals, not dates
AlphaEvolve, WORKBank, agent economies and SE 3.0
9. A verifiable evaluator turns generation into search
10. Algorithm success does not prove full-SDLC autonomy
DeepMind reports real results and a clear applicability condition
Data-centre, chip-design and AI-training optimisation
New matrix-multiplication algorithms
Evaluation must be objective and automated
Ambiguous product goals are not scored this way
11. Automation capability and worker desire: separate axes
12. Agent economies differ by origin and permeability
13. 03. Scenario: intent becomes the interface
A human states the goal; a machine searches the solution space
14. AI-native shifts focus from code to intent
15. Measure durable goals before choosing tools
16. Safe delegation needs a runtime
17. Agents span the SDLC without transferring accountability
18. Agent-written code requires scalable verification
The constraint is proving quality, not generating code
Evaluators must test functionality and risk
Review cost belongs in total task cost
Rising rework falsifies the productivity promise
Safe rollback limits the blast radius
19. 04. Metrics connect agents to outcomes
Acceptance rate helps, but does not show whether the system improved
20. Agentic development needs multiple metrics
21. A pilot needs a stop condition before it starts
Otherwise local acceleration is easy to mistake for impact
A baseline on comparable tasks
Task success and human intervention
Total cost: compute + review + rework
Escaped defects and delivery outcome
22. One signal can lead to different futures
23. By 2030, verifiable systems beat autonomous ones
What evidence must confirm or falsify
Intent becomes more important than typing code
Evaluators become part of development architecture
The platform defines the agent autonomy boundary
Outcomes and rework outrank generation volume
24. Observations are separated from the author's scenarios
Full links are in the speaker notes
DeepMind: AlphaEvolve · Virtual Agent Economies
Stanford: Future of Work with AI Agents
SE 3.0: AI-native Software Engineering
Google Research · DORA: developer goals and metrics
25. Thank you!
polomodov.tech
All slides and links — in the Telegram channel
Alexander Polomodov, Technical Director & Fellow, large fintech
@book_cube