Applied research,
documented in public.

Italian schools went from implicit ban to explicit mandate in twelve months. Whoever builds for education must design governance first — deployer accountability, transparency, AI Act — and the feature second. The reverse order does not pass selection.

Local is no longer a compromise, it is a fleet tier. The operational question of 2026 is not 'open vs closed' but what share of traffic deserves the frontier API. For many enterprise clients the answer is below 20%.

When models converge, orchestration is the product. More agents does not mean more intelligence: it means more coordination to pay for. The value lies in the router assigning the right task to the right model at the right cost.

Skills are to coding agents what pure functions are to code: composable, testable, versioned. A pattern that repeats three times must be extracted. Monolithic prompt engineering gives way to behavior libraries.

A compact, fast model that routes to a specialist outperforms an extended model operating as a soloist. Model tiering — Haiku for routing, Sonnet for execution — is the reference pattern for production workloads.

Skills are not a developer-only artifact. They define a new format for corporate procedures: readable by an operator, executable by an agent, versioned as code. Every repetitive internal process is a skill awaiting formalization.

Concealing the agent's decision process no longer constitutes a product advantage: it generates trust debt. The enterprise user requires understanding of decisions, not mere execution. Exposed routing defines a new UX paradigm for AI tools.

Agent memory does not coincide with a passively populated vector database. It consists of deliberate writing: rules, gotchas, decisions, code breadcrumbs. An agent that identifies what not to remember outperforms one that memorizes everything.

The open vs closed debate is superseded by the hybrid routing paradigm. High-volume repetitive workloads reside on-prem or self-hosted; deep reasoning remains on frontier APIs. The relevant question is where to draw the line, not which model dominates.

Agent parallelization produces value only in the presence of genuinely independent tasks. Naive file-level locking proves inadequate; semantic orchestration based on the functional domains involved is required.

AI does not eliminate the craft, it relocates it. The next-generation professional defines problems better, structures context better, verifies output better. AI amplifies design thinking, not improvisation.

An agent writing without peer review produces mediocre code. A team of agents in mutual peer review produces merge-ready code. Peer review has become a system function, no longer an exclusively human role.

Without persistent shared memory, every AI session restarts from zero. With it, the project acquires historical memory and agents cease repeating already-solved errors. The Second Brain defines the threshold between prototyping and production.

AI proves ineffective at rewriting from scratch. It excels at removing dead code in the presence of a solid test harness. The lever is test coverage, not the model.

Many AI optimizations reside not in the model but in tool configuration. Project Groups have existed for months but adoption remains limited. Systematic reading of changelogs, including secondary ones, represents an undervalued competitive advantage.

A product validated internally is not necessarily validated for market. What appears obvious to the team is the first request from beta testers. The beta serves to unlearn one's own implicit competence.

Loyalty to a single provider constitutes a luxury, cost-aware routing an operational necessity. Selecting the appropriate model for each individual task exceeds in relevance the selection of the best model overall.

For Italian enterprise technical language, the open vs closed gap remains significant. On sensitive projects, a commercial LLM with a clear DPA outperforms immature self-hosting.

As volume increases, visual review fails. Evals constitute the thermometer of agentic quality. They are not an optional lab activity, but a standard production practice.

The data had been available for months, without analysis. AI did not generate new information: it rendered legible what was already present. In some contexts the value of AI does not reside in the answer, but in the formulation of the question.

The interface conditions the workflow to a greater extent than expected. Assigning each agent a distinctive visual representation modifies the mode of work delegation. Usability elements significantly affect real productivity.

The chat interface was designed for conversational turn-taking. Agentic coding is not conversation: it is multi-thread flow. The emergence of alternative interfaces is a distinguishing trait of 2026.

Simple tools surpass sophisticated tools when they support a real workflow. Markdown, folders, an agent that reads and writes: a solution less complex than it appears, still underutilized across many enterprise contexts.