AI4SDLC: What I Would Do Differently

Alexander Polomodov expands his short Deep Tech Night talk into a retrospective on adopting AI in software development. The organizing question is what to rent, adapt, and own. Existing models and agent loops reduce the burden of maintaining a stack; context and adapters connect them to the company’s environment. Permissions, access rules, and evidence of quality remain organizational responsibilities.

Asked about a team of twenty to fifty people, the speaker argues that a smaller organization can often assemble a process from existing products more quickly. A large company faces modified platform tools, scattered knowledge, approval processes, and inconsistent team practices. A capable model alone does not remove these constraints. It needs to be evaluated together with its harness, context, and available actions.

A separate section examines permissions and tools. Agents need limited authority, clear errors, concise responses, and final-state verification. MCP does not replace this work, while repeated human approvals can become a formality. Examples include verbose Figma responses, mechanical OpenAPI wrappers, and skills whose usefulness is never evaluated on actual tasks.

The closing discussion connects quality with cost. A successful demo does not establish reliability: an assistant may work on one project and fail on another. Traces support diagnosis, while real tasks provide repeatable checks for changes. Dividing engineering work into verifiable units makes it possible to compare models, try cheaper execution, and separate planning from operations on internal data. Custom adaptations need reassessment as existing products improve.