#aiinsoftwaredevelopment — Public Fediverse posts
Live and recent posts from across the Fediverse tagged #aiinsoftwaredevelopment, aggregated by home.social.
-
👻 Monster anti-pattern of 2026: pretending GenAI can magically speed up Product Backlog Refinement. Several voices suggest it can write the User Stories, Acceptance Criteria, even do estimation. The trouble: refinement isn't a writing task, it's how the team builds shared understanding. Hand the writing to a machine and you skip the conversation that was the whole point.
If your team has ever handed refinement to a tool to save time, what did you actually lose?
-
🔍 I have a ruleset for Claude around duplicate code. It confidently reported it had checked for duplicates on a 1500-line change. It hadn't, I found one function duplicated three times, and Claude eventually admitted to five.
Skill loss happens wherever we hand a task to GenAI, and if it's right 95% of the time, we still need the skill to catch the other 5%.
Make deliberate choices about what to abandon and what to keep sharp. Practice the skills you intend to maintain. Make it a choice, not an accident.
-
🧪 More and more products are clearly built with GenAI: features that don't fit together, some solving no real problem.
Even when building is cheap, building the wrong thing is still expensive.
Before GenAI the bottleneck was Quality Assurance. GenAI doesn't fix that, it just pushes more code upstream of the same bottleneck. Classic Theory of Constraints.
GenAI speeds up the mechanics of Discovery, which means we need to spend MORE time thinking, not less. Ask each week: what did we learn that changed our plan? If nothing, it's theatre.
-
I spent a week warning people about the risks of LLM code generation. Then I fell into the same trap.
I automated my course-prep emails, the repetitive, error-prone kind. Claude Code copied, tagged, uploaded, and I carefully reviewed and sent. Later I realized it had sent them only to me, and one had the wrong date anyway.
Making mistakes faster is not winning. The automation didn't remove the failure mode, it just got me there sooner. The review step you're counting on is doing more work than you think.
https://agilepainrelief.com/newsletter-subscribe/?utm_source=mastodon&utm_campaign=archive-reshare
-
⏱️ Shortly after ChatGPT 4o, people talked about doing Backlog Refinement in minutes with GenAI. Most have realized that's a bad idea.
Sprint Planning isn't about a tidy Sprint Backlog. It's the team building a shared understanding of the Sprint Goal, capacity, and how they'll get there. A faster event doesn't build that understanding, it skips it.
So don't ask how GenAI makes the event quicker. Ask how it helps the team spot what gets missed and go deeper. Stress-test the plan: yes. Write your Sprint Goal for you: no.
GenAI amplifies what you're already doing, good and bad.
-
🤖 Tools keep promising to automate the parts of Scrum that are entirely about human connection. It's worth naming why they miss.
One vendor replaces the Daily Scrum with scheduled text and video reports for admins. But the Daily Scrum exists to inspect progress toward the Sprint Goal, surface impediments, and plan the day together. An automated tool serves none of that. It has nothing to do with Scrum, except that it stole the language.
Same story for Retrospectives on autopilot from survey submissions. Grouping exercises exist to start a conversation about how problems connect; hand that to AI and you've replaced the human insight that was the whole point.
Data collection and note-taking can help. Replacing the conversation doesn't.
-
🏋️ How I explain Technical Debt to non-technical people: imagine your developers in a large gym. Time them running across and back. Now scatter chairs across the middle at random. They're slower. Add more, and eventually they have to crawl.
The chairs are the technical debt; the extra time is the interest, paid on every trip. AI-generated code has become a major new source of those chairs.
How many has your team quietly added this year?
-
📉 Canada has had a productivity problem: 88% as productive as the US per hour in 1984, 71% by 2022. The new AI strategy promises a 3% GDP boost by lifting business AI adoption from 12% to 60% by 2034.
But adoption of AI doesn't make an organization more productive. Dropping it onto a broken process automates the mess, and an adoption target invites AI washing.
To lift it, treat each organization as a system and fix the bottleneck.
-
🤖 Good news for ScrumMasters: you're safe.
I went looking for useful Claude skills for ScrumMasters, Product Owners and Agile Coaches. What's out there: run the Daily Scrum as a progress summary (it even had a token budget), plan Sprints from capacity and velocity, run Retrospectives where the LLM does all the thinking, turn BDD (a conversation tool) into a rigid spec generator.
Automating the output, skipping the thinking.
-
🏥 Use GenAI to speed up ER triage, but if the hospital is at 100% bed capacity, faster triage barely helps. The beds are the bottleneck.
Humber River Hospital's command centre works because a cross-disciplinary group watches the whole flow. GE brands it "AI", but that pattern-finding predates ChatGPT. Improving the flow made it effective, not AI.
Study the flow before you drop GenAI on a bottleneck.
-
🇨🇦 Canada's new AI strategy promises a 3% GDP boost by raising business adoption of GenAI from 12% to 60% by 2034.
My worry: adoption of GenAI doesn't make an organization more productive. Dropping it onto a broken process just automates the mess. Measure success by adoption rate and orgs will claim they're "doing AI" to hit the number.
To improve productivity, treat each organization as a system and eliminate the bottleneck.
-
🏷️ AI washing: GE promotes Humber River Hospital's command centre as an AI system, but the AI isn't the central feature.
The real value is a cross-disciplinary group paying constant attention to how the hospital works and fixing constraints as they appear. AI helps only because they've studied the flow and set metrics that matter.
The irony? That pattern-finding predates ChatGPT. Flow made the hospital effective, not AI.
-
📉 Losing the skills you stop practising has a name: Deskilling.
In radiology, AI detects abnormalities faster and never tires, but too much reliance leaves the radiologist unable to detect on their own - worse for new doctors who never built the understanding.
Pilots hit this with autopilot and responded with more training, not less. Software is walking the same cliff, but cutting practice instead.
-
🃏 GenAI code quality starts with how models are trained. Training rewards pass/fail: does it compile and pass the tests? There's no reward for code that's easy to maintain. Worse, 0-1 grading rewards guessing over admitting uncertainty, so models learn to "bluff" with overconfident guesses.
OX Security likens this to an army of junior developers: fast, functional, lacking architectural judgment. TDD helps, but solves the wrong problem.
-
🚲 A reader pointed out that no LLM can recommend a trustworthy bike shop. True. And no LLM can do a bike fitting either.
I walked in expecting to buy a $2500 gravel bike. The fitter asked question after question, then steered me to a cheaper one: heavier parts that'll actually last longer. His advice: ride a couple thousand km, see what needs changing, then pay for better.
Embodied, contextual expertise still wins.
-
I shipped a side project with every receipt image publicly exposed. I teach this for a living, I was careful, and I still missed it.
The people being pushed to ship faster with GenAI aren't careless. They're careful people under pressure, with tools that make mistakes easy to create and hard to spot.
Surviving the AI Tsunami is about leading a team that uses GenAI well. Six weeks, real problems from your team. Starts June 29.
-
A research finding that's easy to miss: GenAI can make people more productive and less motivated at the same time.
More output, less ownership of it. For a Scrum team, motivation and shared understanding aren't nice-to-haves, they're the engine. Trade them for raw speed and it's a poor deal.
A third full: useful for the dull work, oversold on the rest. Protect the conversations that keep people invested.
-
🎲 Mandates are landing on desks. Productivity claims are getting louder. Almost no one is bringing the actual research to the conversation.
Three minutes from now you could be the rare person in your org who knows how GenAI actually fails, instead of deciding on vibes.
Test your knowledge of AI Failure Modes in Development Teams.
https://agilepainrelief.com/quiz/ai-failure-modes/?utm_source=mastodon&utm_campaign=archive-reshare
-
🐢 "It feels faster" is not the same as "it is faster."
In one study, seasoned devs using GenAI on familiar code came out 19% slower than without it, while certain they were speeding up. Self-reported productivity and measured throughput barely correlate.
A third full: real value, well short of the hype. Measure the real thing, Cycle Time and Throughput, before you trust the feel.
-
🥃 Most of the AI conversation gives you two options. The glass is overflowing (AI replaces engineers next year) or empty (all hype, avoid it).
I'd put it at about a third full. GenAI genuinely helps with drudge work, and it'll happily turn a 15-minute task into a 3-hour rabbit hole. Both true.
My course, Surviving the AI Tsunami, helps ScrumMasters and POs tell the real third from the noise. Six weeks, evidence-based. Starts June 29.
-
🧠 Microsoft researchers (2025) found something uncomfortable: the more confident an LLM sounds, the less critical thinking people apply to what it says.
Confidence is the trap, and only one of the failure modes quietly shaping how teams build software right now.
How well do you know the AI Failure Modes hitting Development Teams? Test yourself.
https://agilepainrelief.com/quiz/ai-failure-modes/?utm_source=mastodon&utm_campaign=archive-reshare
-
GenAI has one job: produce a plausible response to your prompt.
Not accurate. Not correct. Plausible.
Plausible is a low bar, and it's the bar the technology is actually optimised for. Every other property is a guardrail or a marketing claim.
Test your knowledge of AI Failure Modes in Development Teams.
#AIinSoftwareDevelopment #BuildInPublic
https://agilepainrelief.com/quiz/ai-failure-modes/?utm_source=mastodon&utm_campaign=archive-reshare
-
Ask an LLM why it made a mistake. It does not replay an internal reasoning trace. There isn't one.
It invents a plausible explanation in real time, using the same process that produced the original mistake.
Confidence in the answer is not evidence about the answer.
How many AI Failure Modes in Development Teams can you identify? Test yourself.
#AIinSoftwareDevelopment #BuildInPublic
https://agilepainrelief.com/quiz/ai-failure-modes/?utm_source=mastodon&utm_campaign=archive-reshare
-
AI code generators promise speed. The promise leaves out the part where the bill arrives later.
GitClear's data was the first big public signal: copy-paste exceeded moved lines, churn climbed. Several studies since show the same pattern. Code is cheap to write, expensive to maintain.
-
Microsoft 2025 study on knowledge workers and GenAI: the more confident the AI sounded, the less critical thinking users applied.
Interfaces are designed to sound confident. Users are offloading judgment to a system optimised to convince them, not to be right.
Think you know the AI Failure Modes in Development Teams? Test yourself.
#AIinSoftwareDevelopment #BuildInPublic
https://agilepainrelief.com/quiz/ai-failure-modes/?utm_source=mastodon&utm_campaign=archive-reshare
-
Security defects in AI-authored code: roughly 1.57x the rate of human-authored.
The models trained on whatever code was on the internet. Vulnerabilities and all. They learned the average, not the secure subset.
More AI code = more AppSec load. Nobody put that on the board.
Test your knowledge of AI Failure Modes in Development Teams.
#AIinSoftwareDevelopment #BuildInPublic
https://agilepainrelief.com/quiz/ai-failure-modes/?utm_source=mastodon&utm_campaign=archive-reshare
-
Heavy AI-adopting teams: change failure rates up ~30%. Incidents per pull request up 23.5%.
Speeding up code generation didn't speed up delivery, it just moved the queue downstream into review, testing, and incidents. Theory of Constraints called this out decades ago.
How much of this is on your radar?
Think you can spot the AI Failure Modes in Development Teams? Test yourself.
#AIinSoftwareDevelopment #BuildInPublic
https://agilepainrelief.com/quiz/ai-failure-modes/?utm_source=mastodon&utm_campaign=archive-reshare
-
Cognitive Complexity in agent-assisted repositories has been observed to rise 39%. Static analysis warnings rose with it.
Code that's harder to read is harder to change. Slower future work, more incidents. The bill arrives later, not in the sprint where AI sped you up.
How well do you know the AI Failure Modes hitting Development Teams? Test yourself.
#AIinSoftwareDevelopment #BuildInPublic
https://agilepainrelief.com/quiz/ai-failure-modes/?utm_source=mastodon&utm_campaign=archive-reshare
-
AI agents promised faster delivery. They also delivered 39% more Cognitive Complexity and 30% more static analysis warnings.
A reinforcing loop: more AI code -> complexity rises -> debt accumulates -> velocity drops -> teams write even more AI code to compensate.
Speed gains? Gone within months. Not a tooling problem, a systems problem.
-
AI-generated code: about 1.7x more issues per pull request than human code. At the 90th percentile of teams, that doubles.
Code is cheap to generate. Reviewing and refactoring it is not. The bottleneck didn't move, it just got buried under more PRs.
Test your knowledge of AI Failure Modes in Development Teams. Twelve questions, three minutes.
#AIinSoftwareDevelopment #BuildInPublic
https://agilepainrelief.com/quiz/ai-failure-modes/?utm_source=mastodon&utm_campaign=archive-reshare