Last month we had another meeting for our user group, and I still have great feelings about it, feel it was one of the best in months even though I don’t know why I feel that. Maybe it was how the speakers have been selected – I took those, who didn’t make the cut for CzechDreamin and somehow their topics mostly complement each other and it makes an interesting mix. Put a lot of people registered and actually attending, being in the new Actum‚s offices and much more, it was such a great event.

But the real question is what I took away, besides the great chats, being able to speak with HR person who sees the market from outside, meeting bunch (like a lot) of new people I don’t remember seeing before and hearing a great feedback about how welcoming the community is and how even the new joiners weren’t afraid to jump into the debate and ask questions. Wow!

You can find the recordings (and second part) on Youtube and I shared the presentations on Google Drive.

Dina McLaughlin – The Platform Has Moved on – Have You? Pivoting in the Age of Agentic AI

What a great opener for the whole evening. Dina nicely summarized what happened on the market/in the industry, how it changed her life, how it is all quicker. At the same time we are moving from detesministic to probabilistic systems, where we cannot really trust or predict the output. Our roles will be more about orchestrations, reviewing and guiding, the human-in-the-loop critical due to context, creativity and bias awareness.

AI is already changing day-to-day work – for example junior devs asking AI instead of seniors, which is great and bad at the same time (cannot find the article where I found how crucial it will be to keep the human interaction and experience in the loop as well).

• The pace of change is no longer linear — it’s exponential and hard to keep up with.
• Consultants face the hardest shift → need to become transformation advisors, not just implementers.
• Businesses are confused about AI (what to automate vs what needs AI).
• Rise of autonomous agents + voice interfaces → interacting with systems without UI.
• Strong emphasis on continuous upskilling – learning is now lifelong and non-negotiable.
• Discussion around middle management potentially shrinking (AI replacing coordination roles).

Michal Verner – The Developer Who Wasn’t There: Autonomous Salesforce Development with Claude Code

👉 Can an agent take a user story and deliver a full feature (code → tests → UI → PR) autonomously?

• AI can orchestrate the full pipeline, but only with heavy upfront setup (skills, rules, MD file, tooling) – and I had to laugh here, because it can generate it all but at the same time it expects it from you and you can get it from Reddit or other places, so now we will spend time googling skills and trying the best one just so the AI can generate the things better and quicker at the end (or maybe not)
• I need to learn new words, to fully understand what it is doing – Unravelling, really?

• The cloud.md file is the real “brain”: acts as instruction manual + standards, must be continuously refined, you “earn it” over time through iteration
• quality is inconsistent – sometimes over-engineered (too many guards, extra code), requires review or additional agents for validation
• Interesting shift in effort:
  👉 Today: time spent coding
  👉 Future: time spent designing instructions, pipelines, guardrails
• The agent doesn’t “understand everything” – it searches relevant files on demand, not full context; surprisingly effective even on large codebases
• Dev behavior shift: “I’m not really writing much code these days” → developers moving toward guiding, nudging, reviewing
• it was interesting to me that those agents took the whole presentation (almost 30 minutes) to generate and deploy the code, I always have a feeling the „animation“ of how it write the text is just to slow us down and it can generate the whole code base in a blink of eye

💬 Overall feeling in the room

• Curiosity → “does this actually work?”
• Surprise → “okay, it kinda does…”
• Skepticism → “but is this maintainable / safe?”
• Humor → jokes about “agent replacing husband” / rubber duck debugging
• Realization → this is powerful, but not plug-and-play

Aleš Remta – From 75% Coverage to Sustainable Apex

• Sustainable Apex = long‑term velocity, not short‑term hacks
• Good code must be understandable, changeable, reliable, and modular – clear boundaries and contracts reduce cognitive load and prevent ripple‑effect bugs when modifying code
• Tests are not for coverage – they’re for proving requirements
• Tests are the best design tool – Writing tests first forces better API design, clearer interfaces, and more modular architecture
• Writing tests last leads to brittle, implementation‑cementing tests – When tests are written after the code, they often mirror the implementation too closely, making refactoring painful and discouraging change
• Test‑first exposes requirement gaps early – Writing tests before implementation reveals missing or unclear requirements before code is written, reducing rework and surprises at the end of development
• Use wrapper objects (e.g., DiscountRequest) to stabilize interfaces – Passing primitives or SObjects directly leads to exploding method signatures. A wrapper object keeps the interface stable even as requirements evolve .
• Guiding principles: be intentional, strict with API contracts, write more tests, and write them earlier

Dominik Hlaváč – MCP Servers: Elevating Salesforce Development

• Why MCP matters: AI needs context, and MCP provides it – compares MCP to a “USB cable for AI” that lets the model see your files, metadata, and org context – something base LLMs cannot do
• MCP architecture: client ↔ MCP server ↔ tools/resources – The server exposes tools (APIs, local files, custom logic) that the AI can call. The AI “remembers” available tools and uses them to act.
• Salesforce DX MCP server wraps the CLI into tool groups – It exposes org management, metadata operations, test execution, LWC tooling, and DevOps capabilities
• Ecosystem is exploding: Figma, GitHub, Atlassian, many more – MCP servers now exist for design, code, documentation, and workflow tools. The official registry is still in preview.
• Real use cases: dependency mapping, scratch‑org prototyping, impact analysis – MCP can automate org introspection and selective deployments: “List the org and see the whole trajectory of some attribute…”
• Limitations today: models can be lazy, inefficient, or loop endlessly – who knows how many API tools your „simple“ query will take at the end?
• Security & pricing concerns: API safer than chat, local models possible, MCP likely to become paid – “Running the API is much safer.”

And the group discusses future pricing, API consumption, and guardrails. 9pm and people are slowly heading out of the building, 10pm and we finally closing the day. Thank you all!