My personal thoughts ...
Marketing Hype or Reality?
I have been reading and re-reading the highly recommended Agentic SDLC Handbook and a growing collection of books on AI and software engineering. It all sounds like a timewarp into the future. Yet when I look beyond the "Hello World" examples and polished demos, I still struggle to find many real-world examples that do not come with gazillions of lines of generated code and infrastructure that would take a lifetime to review. After all, lines of code do not equal value, and they certainly do not attract investment.
OK, yes, we could ask agents to validate the output of what the agents generated ... but how can I trust the validation agents? Trust is earned and I, for one, am not in a place where I blindly trust them.
So, while I am not yet ready to bet my entire future on what Microsoft calls Phase 2 (Agent-first) and Phase 3 (Multi-agent) adoption, I have not given up. The agentic future is coming, and it is arriving much faster than many of us anticipated in Where Will We Be in Three Years?. We are already experimenting with Phase 2 activities and are actively exploring Phase 3 possibilities.
GitHub Copilot Usage Metrics moves the behind-the-scenes activities, cost, and value to a transparent forefront. It also positions usage in the above mentioned phases.
That said, I am completely sold on Phase 1.
Phase 1: Value First
Recently, I picked up a support ticket with a seemingly simple request: "Create a list of all Azure DevOps projects in our organisation, including their names, globally unique identifiers (GUIDs), and whether identities X and Y had read access. The first part was easy. The second requires a mind-numbing journey through the Azure DevOps REST Application Programming Interface (API) to figure our identities and access."
Daniel, if you are reading this, please share your Azure DevOps REST API experiences with identities in the comments.
So I asked Agent Ubuntu for help.
Before my wife had finished watching Wheel of Fortune, I had a PowerShell reference script, the completed report, and a closed support ticket. My effort was effectively zero. The cost? 242 AI credits ... a whopping $2.32.
WOW.
That is what I call value. Not flashy demos. Not mountains of generated code. Just a real problem solved in minutes, for much less than the cost of a cup of Mocha.