Coding with AI: a guided starting point

All English wording below is an editorial translation/paraphrase of Italian captions; nothing is a verbatim quotation.

Salvatore Sanfilippo's Italian-language playlist follows coding agents through real systems work rather than presenting one fixed workflow. Its recurring subjects are architecture, specifications, model selection, behavioral testing, local inference, the changing value of implementation work, and the human cost of supervising more work than one person can understand.

The practical thread is expert supervision: agents can generate large implementations quickly, but the human still supplies constraints, representative experiments, product judgment, and the decision about which evidence is enough.

Key themes

• Automation changes where value concentrates without making coding trivial. Sanfilippo separates implementation effort from the conceptual choices that determine what software should be, then argues that models automate part of a programming culture people built rather than erase it (00:00:57–00:03:03(opens in a new tab), 00:10:06–00:12:12(opens in a new tab)).
• Review shifts toward ideas, architecture, and behavior. Sanfilippo argues for spending less time on exhaustive line reading and more on constraints, targeted design questions, manual QA, and model-assisted testing (00:00:01–00:03:55(opens in a new tab)).
• Specifications complement source code rather than replacing it. Mature implementations retain edge cases and interactions that natural language loses, while curated related programs can teach agents reusable implementation patterns (00:03:20–00:05:34(opens in a new tab)).
• Model evaluation belongs on representative work. In one systems-programming test, a passing suite hid ineffective complexity and semantic breakage; Sanfilippo argues that benchmark gains can overstate practical gains (00:12:09–00:15:13(opens in a new tab)).
• A persuasive explanation is not independent evidence. He describes challenging a strong model's plausible diagnosis with observed failure behavior before it reconsidered the conclusion (00:05:44–00:09:03(opens in a new tab)).
• Behavioral QA can explore states that fixed suites miss. Sanfilippo proposes durable QA missions for bounded agents, followed by a separate confirmation pass before a suspected failure becomes an issue (00:03:15–00:09:23(opens in a new tab)).
• Human attention is part of the system. Parallel and asynchronous agents can fragment the developer's working model, extend the workday, and create chronic supervision fatigue (00:04:03–00:09:48(opens in a new tab)).

Where to start

1. Implementation effort, software ideas, and coding agents(opens in a new tab) for the playlist's clearest argument about what automation changes without trivializing programming.
2. Check the ideas, not every generated line(opens in a new tab) for its sharpest review argument.
3. Why specifications do not replace code(opens in a new tab) for a concrete research-to-reference-to-implementation workflow.
4. Automatic programming fatigue(opens in a new tab) for the attention and wellbeing limits of parallel agent work.
5. When a convincing model diagnosis is wrong(opens in a new tab) for reciprocal review and the need for behavioral evidence.
6. Testing alternative models on real systems work(opens in a new tab) for the gap between benchmarks, passing tests, and useful implementation.
7. When local inference makes sense(opens in a new tab) for an end-to-end view of quality, prefill, generation, privacy, control, and hardware use.

The August 16 standalone talk The Qualities of Tomorrow’s Programmers(opens in a new tab) turns the same shift in developer responsibility into practical advice on scope, system understanding, testability, reversibility, consolidation, and communication.

For the full risk-scaled acceptance workflow, read Make the Agent Prove It, or watch the full Coding with AI playlist(opens in a new tab).