DField Q2 2026 roundup · what shifted, what we shipped, what is broken

Q2 2026 is the quarter the demo stopped selling. In April you could still close a deal with a 90-second video and a slick agent walk-through. By the end of June, CTOs are asking for the bill: monthly LLM cost, p95 latency, eval pass-rate, retry-cost budget. If you cannot put numbers on the table, the next RFP shortlist does not include you.

This roundup goes in five parts: what is trending in AI / agentic, what we shipped this quarter, what the Hungarian-market reads like up close, what is broken industry-wide (the strong-opinion section), and what we are watching for Q3.

Trending in AI / agentic

• Anthropic prompt caching crossed from beta into GA. Pricing stable, rate limits up, the system-cache plus tool-cache plus Files-API stack delivers a real 60-80% cost cut on a typical production agent.
• MCP (Model Context Protocol) became the default tool surface. Stripe, Linear, Notion, GitHub all shipped first-party servers. Bespoke function-calling glue has about two more quarters before it is technical debt.
• The agentic SDKs converged. Anthropic Agent SDK, OpenAI Agents SDK, Google ADK all settled on a recognisable tool-plus-state-plus-memory shape. Portability between them is now realistic instead of aspirational.
• On-device LLMs cleared the 'good enough for high-frequency features' bar. Gemini Nano and Apple Intelligence Foundation models hit roughly 80-85% acceptable-output on Hungarian summarisation; cloud fallback is still required.
• Evidence-based AI development reached the consulting layer. 'Eval as code', 'retry budget', 'cost guardrail' are now in CTO vocabulary. That changes how the buying side asks questions.
• On the security side, agentic prompt injection and MCP-server lateral movement are real, documented attack vectors. If you are shipping agents and not auditing your tool catalogue, you are wearing the consequences.

What we shipped this quarter

• Eleven client engagements, four of them production agent rollouts. One Hungarian B2B SaaS: monthly LLM bill from 980 EUR to 220 EUR after two days of prompt-cache work. Real before/after, real invoice.
• A SZEP card 2.0 sandbox integration for a webshop platform vendor: a normalisation layer over three issuing banks (OTP, MKB, K&H) so they only learn one shape.
• Two MCP servers for in-house developer agents. What used to be 800 lines of bespoke function-calling glue collapsed to about 200 lines plus a standard auth flow.
• Long-form posts: prompt-cache wins, MCP for Hungarian engineering teams, evidence-based agentic patterns, NIS2 minimum checklist, SZEP 2.0 checkout, NAV Online Invoice v3, Hungarian webshop CWV.
• Two sales pitches rebuilt around hard numbers · 'what does the CFO pay if this becomes a penalty system, what does it pay if it stays a support assistant'. Conversion came in at 38%.
• Studio infrastructure stayed on the monorepo plus Bun plus Drizzle stack. The tooling investment paid back; nothing in there had to be rebuilt this quarter.

Hungarian market pulse

• SZEP card 2.0 sandbox is live at all three issuers. Live launch is autumn. Anyone not in sandbox now is going to feel the deadline. Spec differences across the three banks are small but real, so a normalisation layer is mandatory.
• NIS2 in Hungary: roughly 70% of SaaS vendors are still reading the law. The first audits ran. The most common gap is not the policy doc, it is the incident-response practice and supplier-contract amendments.
• MNB instant payments 2.0 final spec landed. QR payments and request-to-pay land in 2027, but the webshop integration design has to start now.
• NAV Online Invoice v3 is mandatory in early Q3. Most Hungarian KKV vendors are finishing regression testing. If you are still scoping, you have about six weeks left.
• Consolidation in the Hungarian AI consulting layer. Last year's wave of 'I am also an AI consultant' practitioners is folding into actual delivery shops. Healthy.
• Engineering market: senior AI integration plus Hungarian language is premium. Rates went up 18-22% versus Q1. Hard to find, expensive when found.

What is broken · strong-opinion section

Three things broke at industry level in Q2 2026, and buyers should care.

• Roughly 60-70% of the consulting tier is still selling with demos, not with evidence. 'We will build you an AI agent' in 2026 is the same play as 'we will build you a chatbot' in 2024 · two years of wasted runway if the CTO does not catch it.
• The MCP story is half-understood by the market. Shipping a server is the easy part. The triad of tool listing, auth flow, and lateral-movement audit only exists at a handful of teams. The next public agent incident comes from an unaudited MCP server. Mark the date.
• Low-code agent platforms (n8n plus LLM, Zapier plus LLM, regional SaaS variants) are a trap on production. They produce a fast POC, then choke the moment you ask for evals, cost guardrails, retry policies. The CFO asks for numbers two months in and there are none.
• Hourly billing in AI consulting is on its last legs. Vendors who win this quarter quote fixed-scope, evidence-table outcomes. Time-and-materials still wins on a few legacy buyers, but the pricing pressure is one-way.
• AI security is on both the CTO and the CISO desk, but the conversation between them is not happening. The CTO ships the MCP server, the CISO is supposed to audit it, and the gap in the middle is where the incident lives.

On the 'not broken' side: Hungarian webshop performance finally moved. Q2 CWV measurements are 12-15% better than the same quarter last year on LCP and INP. Speculation Rules API plus on-demand image optimisation plus edge SSR is doing the work. That one is genuinely a good story.

What we are watching for Q3

• EU AI Act full effect on August 2. The final high-risk classification list. Hungarian SaaS vendors most exposed: HR-tech and fintech.
• SZEP card 2.0 production launch in September. The normalisation layer, the webhook reconciliation and the error-translation layer all need to be ready before then. No room for autumn-discovery work.
• NAV Online Invoice v3 cutover in early Q3. The week after go-live is the real test · monitoring and observability are mandatory, not optional.
• Next.js 16 GA. Turbopack default, 'use cache', React Compiler. The migration is not free, but pushing it into Q4 is technical debt with interest.
• On-device LLM benchmark season. Public results in late July: Gemini Nano, Apple Intelligence, locally fine-tuned Mistral 7B, all on Hungarian content.
• Hungarian fintech cybersecurity audit framework. The MNB-NHH joint guideline draft lands in August, finalises in autumn. Banks should be reading the draft now.

Q2 in one sentence: the quarter pricing stopped being a deck slide and started being a measurement. Demos are out. Receipts are in. We are going to keep shipping the bill-shaped numbers next quarter, the same way.