By September 2026 the industry’s own framing had shifted: the value is in workflows, data rights and human review, not flashy demos. That’s a good moment to take stock of what has actually changed in how software gets built. It’s easy to be either dismissive or breathless about AI’s effect on software development, and both postures avoid the more useful question: what, specifically, is actually different about how software gets built today versus three years ago? A few concrete shifts are worth naming, separate from the marketing around them.
From autocomplete to agents
The first wave of AI coding tools completed lines and functions. The current wave plans and executes multi-step tasks — reading a codebase, writing code across several files, running tests, and iterating on failures with limited supervision. That’s a change in kind, not just speed: the unit of work an AI touches has grown from a line to a feature, and the human’s job has shifted from typing to reviewing and directing.
Standardized plumbing is enabling standardized agents
Protocols like MCP are doing for AI tool access what REST and later GraphQL did for web APIs: turning bespoke integration work into a shared, reusable layer. That standardization is what makes it economical to build agents that reach into real systems, rather than agents that only talk. See our companion piece on what MCP actually is.
The bottleneck has moved, not disappeared
Code generation was rarely the actual constraint on software delivery — requirements clarity, architectural judgment, and review capacity were. AI tooling makes the easy 80% of implementation faster, which means the remaining 20% — the ambiguous requirements, the cross-system trade-offs, the “is this actually the right thing to build” conversations — is now a larger share of where a team’s time goes.
What we’d actually bet on
- Brownfield-aware AI coding practices becoming table stakes, not a specialty (see our piece on brownfield vs. greenfield).
- Agentic workflows handling more internal, reversible work before they’re trusted with customer-facing actions.
- Governance — audit trails, approval gates, rollback paths — becoming a normal line item in AI project scope.
- Model and provider choice being treated as a swappable architectural decision rather than a permanent commitment.
Summary: None of this requires believing AI will write all the code or none of it. It requires paying attention to where the actual mechanism has changed, and building the engineering discipline around it before the incidents force the issue.
#AI #SoftwareDevelopment #AgenticAI #FutureOfWork #SazkoSolutions