[{"data":1,"prerenderedAt":46},["ShallowReactive",2],{"article-en-ai-coding-team-collaboration-three-patterns":3},{"article":4,"related":17,"prevArticle":35,"nextArticle":39,"availableLocales":43},{"id":5,"slug":5,"title":6,"excerpt":7,"content":8,"image":9,"date":10,"published":11,"status":12,"category":13,"keywords":14,"meta_description":15,"og_image":16,"custom_css":16,"created_at":10,"updated_at":10},"ai-coding-team-collaboration-three-patterns","AI coding in teams: three patterns we keep seeing","After working with more than ten development teams over the past six months, we keep seeing the same thing. Individuals getting faster with AI doesn't mean the team is ready to get faster together.","\u003Cp>Over the past six months, we've worked with somewhere over ten development teams.\u003C/p>\n\n\u003Cp>Their tools vary. Some use Cursor, some have started building Claude Code into their daily workflow, and others are experimenting with agent frameworks, MCP or automated workflows of various kinds.\u003C/p>\n\n\u003Cp>On the whole, almost nobody doubts that AI coding helps. Plenty of engineers already use it very well. One person can get a prototype up quickly, clear away repetitive work, and find their footing faster in an unfamiliar technical area.\u003C/p>\n\n\u003Cp>But once we got into how these teams actually work together, we kept running into a few strikingly similar problems. And usually it isn't just one of them. All three show up at once.\u003C/p>\n\n\u003Ch2>Pattern one: individuals got faster, but the team doesn't know how to speed up together\u003C/h2>\n\n\u003Cp>Some engineers are very good with AI. They know how to break down a task, how to write prompts, and how to get AI to search the codebase, fill in tests and refactor. Some can even assemble a fairly complete workflow on their own.\u003C/p>\n\n\u003Cp>But once that ability has to scale from one person to a team, things change.\u003C/p>\n\n\u003Cp>Someone uses AI to ship a feature quickly, but nobody else knows which assumptions it changed. Someone tries a new framework, and the next person to pick up the work struggles to make sense of it. The volume of PRs goes up, but review quality and review time don't keep pace.\u003C/p>\n\n\u003Cp>The end result: everyone is more efficient, yet the team as a whole is fighting fires everywhere.\u003C/p>\n\n\u003Cp>This isn't necessarily AI's fault. Often, AI has simply amplified collaboration problems that were already there but hadn't started to hurt.\u003C/p>\n\n\u003Cp>If requirements were never spelled out, AI can now turn vague requirements into a big chunk of code even faster. If there was never agreement on architecture, everyone can now head off in different directions even faster. If testing and review already depended on a handful of people, those bottlenecks become even more obvious once output speeds up.\u003C/p>\n\n\u003Cp>AI has made writing code faster. But \"what are we building?\", \"what counts as done?\" and \"who owns this decision?\" don't become clear on their own.\u003C/p>\n\n\u003Ch2>Pattern two: everyone says they're more efficient, but estimates haven't shrunk\u003C/h2>\n\n\u003Cp>This one puzzled us at first.\u003C/p>\n\n\u003Cp>The engineers all say they use AI, and they really do write code much faster. Yet when it comes to schedules, estimates and delivery dates, the numbers haven't noticeably improved. Some teams even feel work has become harder to estimate than before.\u003C/p>\n\n\u003Cp>On reflection, it makes sense. A development task has never been only about writing code. It also involves understanding the requirements, the system context, edge cases, data handling, integration, testing, release and operations, plus all the problems nobody saw at the start.\u003C/p>\n\n\u003Cp>AI can quickly give you a version that looks usable, but it won't judge for you whether the logic matches the real business rules. Nor will it know whether a change that looks perfectly reasonable will break a process someone left behind three years ago that is still running today.\u003C/p>\n\n\u003Cp>So it isn't that many teams haven't become more efficient. The gain happens in the \"producing code\" stage, while the confirming, integrating, fixing and risk-owning that follow haven't shrunk along with it. If anything, because more can be built faster, there is even more to confirm afterwards.\u003C/p>\n\n\u003Cp>If a team absorbs AI's speed with its old way of splitting work, its old acceptance process and its old division of responsibilities, what it ends up with may well be not faster delivery, but faster chaos.\u003C/p>\n\n\u003Ch2>Pattern three: we used to blame each other, now we all agree it was the AI\u003C/h2>\n\n\u003Cp>In the past, when something went wrong with the system, the usual routine was for the team to go round checking:\u003C/p>\n\n\u003Cp>Were the requirements unclear?\u003Cbr>\nDid the engineer get it wrong?\u003Cbr>\nDid QA miss it?\u003Cbr>\nOr did the PM leave scenarios out?\u003C/p>\n\n\u003Cp>These days it's sometimes a bit simpler. Everyone pretty much agrees: it's the AI's fault.\u003C/p>\n\n\u003Cp>Of course, that's sometimes said as a joke. But there is a real problem behind it worth paying attention to.\u003C/p>\n\n\u003Cp>AI can help generate code, suggest approaches, fill in documentation and even help analyze errors. But it is not a team member. It doesn't understand the client's real situation, it doesn't know the history behind your system, and it won't take responsibility for a production incident.\u003C/p>\n\n\u003Cp>Quite a few community figures who became well known through AI have said that clients ask them whether they can lower their quotes now that AI helps with development. Their answers tend to be the same: AI won't take the blame for you. It sounds a little like a threat, but what usually lies behind it is that AI hasn't given them any room to cut prices, and their costs may even have gone up.\u003C/p>\n\n\u003Cp>At its core, this is still about what it costs us to deliver a service at a given level of quality, and whether today's team members can still put in enough time and energy to carry that cost.\u003C/p>\n\n\u003Cp>For a manager, the real question is usually not \"was this AI's fault?\" It is: why could something that nobody had understood, verified or challenged make it all the way into the system?\u003C/p>\n\n\u003Ch2>Step back from the tools, and the real issue is collaboration\u003C/h2>\n\n\u003Cp>What really needs addressing is how the team works together.\u003C/p>\n\n\u003Cp>Tools will keep changing and models will keep improving. Today everyone is comparing which agent is strongest; tomorrow a new way of working may come along.\u003C/p>\n\n\u003Cp>But the questions a team actually has to answer are more basic:\u003C/p>\n\n\u003Cul>\n\u003Cli>Which work can AI take on at scale, and which still needs a clear human judgment?\u003C/li>\n\u003Cli>When output speeds up, how should requirements, design, review and testing adjust together?\u003C/li>\n\u003Cli>Does the team genuinely agree on architecture, quality and what counts as done?\u003C/li>\n\u003Cli>When something goes wrong, can we trace back which decision caused it?\u003C/li>\n\u003C/ul>\n\n\u003Ch2>Why we built Teams\u003C/h2>\n\n\u003Cp>That's also why, when we built our own agent framework last year, we deliberately called it \u003Cstrong>Teams\u003C/strong>.\u003C/p>\n\n\u003Cp>It was never designed as a tool that lets AI run off on its own. It was designed as a set of points where people and AI make decisions together. As a task moves through requirements, design, development, testing and release, each stage can carry different assumptions, constraints and responsibilities. AI can get a lot done quickly, but it doesn't automatically know which judgments can be handed over and where a person still needs to step in and confirm.\u003C/p>\n\n\u003Cp>So Teams keeps a number of key approval and review steps, and requires the people responsible to actually take part in the process. On the surface these steps look like extra sign-offs. In practice, they help the team gradually move judgments that used to be scattered across personal experience, verbal agreements and after-the-fact fixes into a place where they can be seen, discussed and owned.\u003C/p>\n\n\u003Cp>After six months of refinement, more than 60% of the decisions in Teams can now be handled entirely by AI.\u003C/p>\n\n\u003Cp>That isn't because we knew from the start what could be automated. It's because, through real use, we gradually got a clear view of the conditions, exceptions, verification methods and boundaries of responsibility, and only then felt confident handing things over.\u003C/p>\n\n\u003Cp>Often the problem isn't that something can't be done; it's that we didn't see it clearly enough at first. Trusting a smooth-looking process too early and pushing work out too fast doesn't necessarily make you more efficient. Sometimes it just hands over the risks you hadn't yet understood along with the work.\u003C/p>\n\n\u003Cimg src=\"/images/blog/teams-human-ai-decision-handoff.jpg\" alt=\"Diagram of Teams moving from joint human and AI decisions, through verification, step by step to autonomous AI decisions\" />\n\n\u003Cp>Teams doesn't hand responsibility to AI all at once. As the conditions that can be verified become clear, decisions shift over one step at a time.\u003C/p>\n\n\u003Cp>What Teams aims for is a transitional design. First, people and AI decide together at the key points. Once the team understands the facts, rules and exceptions well enough, responsibility moves step by step to AI decisions that have been defined, verified and can be traced.\u003C/p>\n\n\u003Cp>That way, AI's autonomy isn't built on a result that happened to go well once. It's built on the team genuinely knowing why it can be trusted, under what conditions, and, when something goes wrong, which step to go back to and fix.\u003C/p>\n\n\u003Ch2>It was never just luck\u003C/h2>\n\n\u003Cp>Some teams say: let AI run for now, keep going if it works, and fix things when problems come up.\u003C/p>\n\n\u003Cp>In the short term, that sometimes works. The problem happens to be a familiar one, the requirements happen not to have many exceptions, the person in charge happens to spot what's wrong, and it gets through.\u003C/p>\n\n\u003Cp>But it makes it easy to put too much trust in a one-off result. What we're looking at may just be one sample that happened to succeed, not a capability we can actually repeat. With good luck, the feature goes live smoothly. With bad luck, the problem gets stuck somewhere between requirements, architecture, testing and boundaries of responsibility, and in the end everyone can tell something is off, but nobody can say clearly who should deal with it.\u003C/p>\n\n\u003Cp>But is it really a matter of luck? Often it isn't.\u003C/p>\n\n\u003Cp>It's simply that the team hasn't yet put the things that always needed confirming into a clear enough process. When responsibility, approval and verification rest on the experience of a few people, AI just makes that risk blow up sooner.\u003C/p>\n\n\u003Cp>Truly stable collaboration with AI isn't about hoping it works out every time. It's about the team knowing where to stop, who should make the call, and, once the call is made, where responsibility goes.\u003C/p>\n\n\u003Ch2>Conclusion\u003C/h2>\n\n\u003Cp>Individuals knowing how to use AI matters, of course.\u003C/p>\n\n\u003Cp>But if a team wants AI to make it genuinely steadier and faster, what it needs to build isn't more prompts, or just more frameworks. It needs to put collaboration, responsibility and judgment back where they belong.\u003C/p>\n\n\u003Cp>Otherwise AI will only help everyone build more, faster.\u003Cbr>\nAnd set things on fire faster too.\u003C/p>","/images/blog/ai-coding-teams-editorial.jpg","2026-08-20T08:00:00+00:00",true,"published","perspective","AI coding, AI collaboration, agent framework, Teams, development team, software development process, decision accountability","Lessons from 10+ teams adopting AI coding: it amplifies collaboration problems, estimates don't always shrink, and Teams shifts decisions to AI step by step.",null,[18,24,29],{"id":19,"slug":19,"title":20,"excerpt":21,"image":22,"date":23,"category":13},"ai-sdlc-transformation-with-teams","What we learned working alongside teams through an AI SDLC transformation","Since late last year, we have worked alongside 8 teams through a full AI SDLC transformation, and 6 of them are now fully up and running. Looking back, the framework itself wasn't what took the time. Taking each project live, one at a time, and adjusting it again and again was.","https://cdn.sqa.tw/articles/ai-sdlc-teams-cover-2026-11-21.jpg","2026-11-21T12:00:00+08:00",{"id":25,"slug":25,"title":26,"excerpt":27,"image":28,"date":23,"category":13},"what-is-fde-forward-deployed-engineer","What is an FDE? On-site engineers deliver systems, FDEs deliver results","How is an FDE (Forward Deployed Engineer) different from an on-site engineer? One delivers a system; the other delivers a result. Drawing on three weeks one of our engineers spent working alongside a trading company's cashier and sales team, we look at how an FDE works and when a business needs one.","https://cdn.sqa.tw/articles/what-is-fde-cover-2026-11-21.jpg",{"id":30,"slug":30,"title":31,"excerpt":32,"image":33,"date":34,"category":13},"ai-development-sdlc-release-verification","AI builds it in two hours, but verification still takes three days","AI can speed up development, but spec sign-off, testing and release don't speed up with it. Drawing on a booking system SaaS we worked with, we explain how we reworked the SDLC to bring full release verification down from two or three days to under two hours, and how the SQA role shifted as a result.","https://cdn.sqa.tw/articles/ai-sdlc-verification-bottleneck-cover.jpg","2026-11-05T12:00:00+08:00",{"slug":36,"title":37,"date":38},"why-we-are-reintroducing-dbam-ai","Why we redesigned SQA.TW","2026-08-17T16:00:00+00:00",{"slug":40,"title":41,"date":42},"sme-guide-to-choosing-ai-agent","With AI agents everywhere, how should small businesses choose?","2026-09-01T15:00:00+00:00",[44,45],"en","zh",1795282791649]