AI4SDLC: what I would do differently
If I knew then what I know now
Slide contents
1. AI4SDLC: what I would do differently
If I knew then what I know now
2. First, I would draw the ownership boundary
Rent · adapt · own
3. Rent what the market improves faster
Models — Frontier capability portfolio · providers directly or through OpenRouter
Agent loop — Leading harnesses: Claude Code; open Codex and OpenCode
Infrastructure — Standard compute and serving · cloud first, vLLM when needed
4. Adapt the moving boundary to your environment
Routing — Smart task-to-model routing with budget optimization
Context — Packaging, memory, and compaction tuned to platform-specific constraints
Adapters — Semantics across models, tools, and runtime
5. Own authority and switch providers
Authority — Tool contracts, access, and policies—where architecture meets security
Evidence — Owned evals and outcome traces
Switchability — Portable replay and a provider-switch drill
6. The model does not define the system
One checkpoint produces different production systems
7. Capability without authority cannot deliver outcomes
Model capability becomes an outcome through authorized action and independent verification
8. Advantage comes from combining the layers correctly
Changes become improvements as they pass through the entire stack
9. Not every improvement requires a new model
Fast and slow loops improve different layers
10. Most improvements ship through the fast loop
Route to another model
Repackage the context
Change a tool description
Tighten policy or gate
11. Slow-loop infrastructure changes the feasible architecture
Precision · memory · fabric · serving
12. At least four ways close the loop
Vertical · big labs · product-first · open-weight ecosystem
13. A vertical closes the loop internally
Product → evidence → behavior → economics → next release
14. Big Labs own the model loop
Research → model → harness → evals inside; compute with AWS
15. Cursor closed its loop; SpaceX added compute
SpaceX acquired Cursor on August 14, 2026
16. Open weights move the assembly point
17. A model learns a particular action world
A compatible schema does not guarantee compatible behavior (well documented by Cursor)
18. The harness covers the next weakness
A mature capability becomes a tool
19. The contract bounds action consequences
Identity · effect · provenance
20. Observing does not imply training rights
Telemetry · evals · training are different modes
21. A convincing trace cannot prove the outcome
Trace
Actions and tools
Errors and recovery
Cost and policy events
Outcome
Environment end state
Hidden executable checks
Repeated-run reliability
22. An improvement must survive replay
Capture task, trace, and final state
Replay in a frozen environment
Change one controlled layer
Run the same independent checks
Release only a measured improvement
23. Turn the SDLC into verifiable episodes
Keep the frontier model where search is required
24. The workload profile determines the serving architecture
Colocate · split phases · disaggregate KV storage
25. Repeatable episodes fit smaller models
The same gate · escalation under uncertainty
26. A frontier model is not needed everywhere
Your own learning loop is always required
Buy capability, not dependency
Own verifiable episodes
Remove repeatable variability
Uncertain → frontier · repeatable → system