Brian Graham calls it the Knowledge Cliff: automating away understanding you need.
GPS replaced taxi drivers’ mental maps. NASA lost rocket expertise to retirement. In software, every LLM decision erodes skill while giving false confidence.
Before automating, ask: Can I evaluate this? Is this core to my craft?
Automate small tasks. Protect ones that build judgment.
https://kb.buildingbetterteams.de/docs/AI-Augmentation/antipattern-knowledge-cliff/
AI agents promised faster delivery. They also delivered 39% more Cognitive Complexity and 30% more static analysis warnings.
Research found a reinforcing cycle: AI generates more code → complexity rises → debt accumulates → velocity drops → teams generate even more AI code to compensate. Initial speed gains? Gone within months.
This isn’t a tooling problem; it’s a systems problem. More code was never the goal.
GenAI is making it easier to switch tool providers. I’m frustrated with our email marketing platform: most features are GUI-only, no API access.
I’ve written code to extract our entire newsletter archive + stats. When the time comes, exporting will be painless.
The irony: our provider keeps adding AI features I don’t need instead of making their product easier to use.
We don’t need more shiny AI features. Ease of use matters more.
AI models are trained to bluff.
Pass/fail training offers no reward for admitting uncertainty. Instead of "I don't know," you get a confident wrong answer. Researchers call it "test-taking mode."
They learned from the internet's code. No grasp of architecture or design patterns. Just next-token prediction dressed up as engineering.
And newer models aren't always better. Some now silently remove safety checks.
https://agilepainrelief.com/blog/genai-code-quality-fundamental-flaws-and-how-bluffing-makes-it-worse/
"I tested it and it works great." The most dangerous sentence in AI usage.
Test an LLM a few times, it mostly works, you assume it works. But failure might show up 1 in 20 tries. You'll never run it enough to know.
Multiply across a team: six people, ten AI tasks per week. A small error rate means several unnoticed mistakes weekly.
Catching them requires what we're short on: domain expertise, critical thinking, time. Never assume the output is correct.
Heavy GenAI users now prefer it over teammates -- it's "more empathetic and non-judgmental."
That should worry us. When people turn to AI instead of colleagues, team cohesion weakens. You don't have a team, just people in the same virtual space.
The fix isn't less AI. It's intentional collaboration: pairing, ensemble work, team autonomy.
The result? Code that looks correct but quietly introduces bugs that surface in production at 2 am.
Irony: LLMs helped me research this article across 15 sources. They're genuinely good at that. They're just not good at writing code that won't bite them later. The tool matters less than how you use it.
Claude Skills is a clever idea with no support for updates, organization, etc.
I just stumbled across https://www.agenstskills.com/ -- is anyone using this?
"Reasoning" models don't reason. They rely on statistical pattern matching, often just replicating patterns from training data. That's not logical deduction.
More effective than models without it? Sure. But let's be honest—it's still a stochastic parrot.
Who in the BDD world is exploring how to use GenAI and BDD together?
I’ve done early experiments, but want to learn from others.
Please boost for visibility.
These are just some tradeoffs we'll explore in my free webinar: "Where is AI Taking Software Development?"
No hype. No AGI promises. Just the practical question: How is our work changing, and what can we do about it?
OReilly has a very good analysis of how AI might play out across all industries in 2026 and beyond: https://www.oreilly.com/radar/what-if-ai-in-2026-and-beyond/
They're not trying make perfect predictions, instead run a range of scenarios. Based on the past few weeks, I think that OpenAI is in trouble.
On the other hand, many of the other scenarios seem plausible, including the cost of inference drops to almost nothing, or uptake of LLMs is limited by energy, or unrest as job losses++. #AIinSoftwareDevelopment
I've spent 10+ years encouraging ScrumMasters to use #SystemsThinking to understand team challenges. The problem? It's hard to remember all the questions. That's where GenAI actually helps. I built a Claude skill to save some effort: instead of memorizing a framework, you tell the AI your story. It asks the follow-up questions that uncover the deeper issues.
The skill includes Systems Thinking questions and Causal Loop Diagrams
#AIinSoftwareDevelopment #BuildInPublic
https://agilepainrelief.com/blog/genai-systems-thinking-team-problems/
As AI coding in chat platforms becomes mainstream, organisational control and security are at risk. Democratised code generation demands strong governance, human oversight, and careful risk management to prevent chaos and vulnerabilities.
Discover more at https://smarterarticles.co.uk/when-code-lives-in-chat-the-dangers-of-democratised-development?pk_campaign=rss-feed
#HumanInTheLoop #AIinSoftwareDevelopment #Security #Governance
These tools can be very helpful, but we need to know our technical stack and problem domain well enough to judge their output.
#AIinSoftwareDevelopment #BuildInPublic
Join me Fri Feb 13, Where is GenAI Taking the Software Industry: https://buff.ly/rP3KJER
I've argued that while GenAI can reduce critical thinking, using it well actually requires more critical thinking. So how can these tools help develop that skill?
For 10+ years, I've encouraged ScrumMasters to use Systems Thinking for deeper problem understanding. But there's a big leap from textbook examples to real work. Enter GenAI. I've created a Claude skill for Systems Thinking basics.
#AIinSoftwareDevelopment #BuildInPublic
https://agilepainrelief.com/blog/genai-systems-thinking-team-problems/
#ClaudeCode automates tasks but made missteps exporting BufferApp data. Asked for LinkedIn→Mastodon→BlueSky scrapers, got 3 separate codebases instead of reusable components. Different CSV formats, wasted tokens running code in chat vs local.
Still useful short-term until Buffer's API arrives. Frees me for deeper work.
Join me Feb 13: Where is GenAI Taking Software https://buff.ly/rP3KJER
#AIinSoftwareDevelopment #BuildInPublic
Downloading Buffer analytics, the internet suggested "let Claude drive Playwright." Not tempting.
The realization: AI automation has ongoing "cognitive tax"—checking results every run. Better: AI writes code once, I work harder upfront but run safely after.
Bonus: more efficient with tokens/electricity. Reduce cognitive tax?
Join Feb 13: https://buff.ly/rP3KJER
#AIinSoftwareDevelopment #BuildInPublic
Agent Psychosis: Are We Going Insane? https://lucumr.pocoo.org/2026/1/18/agent-psychosis/ - I think it should have been called Agents gone wild. It could be a bad movie on Netflix.
Working in small chunks, using BDD/TDD like approaches - I'm getting good results for fewer tokens.
This agent use seems mad. Poor quality; Burning tokens like they're going out of style. They're treating these tools like they're free and that won't last.
