From vibe coding to agentic engineering: what changes for Italian SMEs from 1 June 2026
On 1 June 2026 two things happened — one visible, one cultural. The visible one: all GitHub Copilot plans — Pro, Pro+, Business, Enterprise — moved to usage-based billing via AI Credits, where 1 credit equals $0.01 and consumption is calculated on the actual tokens (input, output, cached) of the chosen model. The cultural one: Andrej Karpathy officially declared the end of vibe coding as the dominant paradigm, labelling it the "prompt-and-pray" phase of 2024–2025, and consecrated agentic engineering as the discipline with which serious companies write AI-first software. For Italian SMEs these two pieces of news converge on a single practical question: how do we turn a predictable AI budget into measurable productivity, without being overwhelmed by consumption costs?
🗓️ What changed on 1 June 2026 (and why it matters)
Until 31 May 2026 GitHub Copilot was sold as a flat-rate subscription: you paid the seat and used premium models as much as you liked. From Monday 1 June the model changed: the base price of the plans remains the same ($10 Pro, $39 Pro+, $19 Business, $39 Enterprise), but it includes only a monthly AI Credits allowance — once that is used up, consumption kicks in. Inline completions and Next Edit suggestions remain free and unlimited; everything else — advanced chat, Copilot Workspace, automated code review, agent mode — costs extra.
The figure that changes the calculation for an SME is something else: Copilot code review also consumes GitHub Actions minutes in addition to AI Credits. This means that every PR opened with automated review now comes with an attached bill. For the first three months (1 June – 1 September 2026) Business and Enterprise plans receive an increased credit allowance as a transition window, but it is a ramp: after that, consumption discipline becomes permanent.
During the same period Microsoft cancelled a large share of its internal Claude Code licences, redirecting teams to Copilot CLI; Uber exhausted its 2026 AI coding budget in the first four months; and Visual Studio Magazine captured developer dissatisfaction ("less service, same base price"). The message for business owners is clear: the "unlimited licences for everyone" model is officially dead.
⚖️ Vibe coding vs agentic engineering: the definition that matters
Karpathy used two memorable phrases. Vibe coding = "I do what I feel like, let's see what comes out": quick prompt, opportunistic generation, a prototype that works as long as it holds. Agentic engineering = "I write a spec, an agent plans it, implements it, tests it, delivers it, under structured supervision". It is not a question of model (Opus 4.8 vs Sonnet 4.6 vs GPT-5.5) — it is a question of workflow.
The practical difference for anyone managing a team: in vibe coding the developer is the prompt engineer; in agentic engineering the developer is the director of a pipeline composed of planning, specialised sub-agents, test execution, error recovery, and a final human code review. The prompt is no longer the deliverable: the deliverable is the repeatable workflow.
IBM, AWS, Google Cloud, Databricks and GitHub, in recent weeks, have all converged on an operational definition of AI agent: a system with an objective, memory, a plan, tool use, bounded autonomy and a human-in-the-loop. SMEs that keep buying AI "as a copilot" without redesigning their workflow will pay the consumption cost without capturing the benefits of orchestration.
- Vibe coding: single prompt, opportunistic output, no tests, no governance.
- Agentic engineering: functional specification, agentic plan, tests run by the agent, human review.
- The value lies not in the model but in the repeatable workflow that uses it.
- The transition requires 1 senior engineer acting as director, not 10 prompt heroes.
🏢 Why the crossroads for Italian SMEs is right now
A typical Italian SME — 10–150 employees, 1–3 in-house developers or an external partner, inherited management systems, custom ERPs, a desire to automate but a fear of getting burned — faces a perfect convergence of pressures. Vibe coding had created an illusory expectation: "with AI everyone can write software". The reality of 2026 says the opposite: AI makes the expert developer 5–10× more productive, it does not replace the director's role. And with consumption-based billing, every vague prompt carries a cost you previously never saw on your bill.
The competitive advantage is precisely here. An SME that adopts right now the agentic engineering approach — with clean repos, prompts written as specifications, budget alerts, mandatory human code review — can ship to production in 8 weeks what a competitor in "wild vibe coding" mode will never finish, because it will crumble under the weight of costs and bugs. The window of advantage is short, because the competition is waking up at the same moment.
Summary of the new pricing model published by GitHub in May 2026.
✅ The 3 signals that your company is ready for agentic engineering
Not every SME is ready to switch overnight to agentic engineering: attempting it without the foundations is the fastest way to burn budget with no results. When I work with a client, I assess three signals. First: does a repo with a working CI pipeline and at least tests for the critical areas exist? Without tests, agents have no feedback loop and produce plausible but unverified code. Second: is there at least one engineer who can act as director, capable of reading generated code, challenging it, and deciding when to stop a workflow that is drifting? Third: does a budget alert exist — even just via a dashboard — that tells you if this month you are about to exceed the expected AI Credits threshold?
If all three signals are present, the transition is an 8-week roadmap. If even one is missing, it is better to start there before touching agentic workflows. Skipping the prerequisites phase is the mistake that leads most companies to say "we tried AI, it doesn't work".
- 01Weeks 1–2: prerequisitesGreen CI, tests on critical areas, AI cost baseline from the previous month.
- 02Weeks 3–4: pilot1 agentic workflow on a high-value task (refactor, audit, migration).
- 03Weeks 5–6: governanceBudget alert, human review policy, cost-per-PR and post-release bug KPIs.
- 04Weeks 7–8: rolloutExpand to 2–3 workflows, train internal directors, write documentation.
⚠️ Mistakes to avoid in the next 90 days
The first mistake is confusing Copilot Pro at $10 with an unlimited AI budget. With the new model, a single inattentive developer can exhaust the monthly allowance in 2 days if they use agent mode on complex tasks without measuring usage. Setting per-user budget alerts immediately is non-negotiable.
The second mistake is leaving prompts in the hands of junior developers without supervision. In agentic mode, a vague prompt is an incorrect functional specification multiplied across N sub-agents. The initial prompt for a workflow must be written collaboratively and reviewed the way a user story is reviewed.
The third mistake is skipping the human code review on generated PRs. Copilot's new automated code review feature consumes GitHub Actions minutes and Credits, but it does not replace a senior's judgement on architectural coherence. The rule that works: agent suggests, human merges.
SME that burns budget
- Copilot sold to the team as "unlimited AI"
- No budget alerts or per-user caps
- Junior developers running agent mode unsupervised
- Vague initial prompts that explode into million-token runs
- PRs generated by agents merged without human review
SME that works
- Pilot on 1 high-value agentic workflow
- Per-user budget alerts and monthly team cap
- Prompts written collaboratively as a functional specification
- 1 senior director approving every critical workflow
- Mandatory human code review before merge
- Monthly KPIs: AI Credits per PR, % test pass, post-release bugs
Frequently asked questions about agentic engineering
Should I abandon GitHub Copilot and switch to Claude Code or Cursor?
No, not automatically. The crossroads is not which tool but which workflow. Copilot remains competitive for inline completions, which under the new model remain free. Credit consumption applies to agent mode, advanced chat and automated code review: that is where competition with Claude Code and Cursor plays out on cost-per-task. Measure 30 days with your own flows before deciding.
What does it actually cost to use agent mode on a real task?
It depends on the model and the length of the run. With a typical agentic workflow on an average refactor you can consume from 50,000 to 500,000 tokens: at the published prices of premium models, we are talking about a few euros per run. Without budget alerts a team of 5 developers can easily reach hundreds of euros in extra charges per month. The point is not the individual cost — it is predictability.
Can my SME start without an in-house senior engineer?
Yes, but you need an external director for the first 90 days. The most common mistake is thinking AI compensates for a lack of seniority: with agentic engineering the opposite is true — the director role is precisely the part AI does not replace. An experienced consultant who sets up workflows, governance and KPIs in the first 90 days costs far less than the budget you would burn in wild vibe coding mode.
Let's talk
If this topic is relevant to you, write to me: comparing notes on code and AI is always time well spent.