EQT Ventures reposted this
If you want to build “Spec” to “Pull Request” software development automation 🤖, what use case would you tackle first with current LLM models at your disposal? 💡 Two key questions come to mind: 1️⃣ - In which type of setup will you already have detailed specs and context fully prepared in the current software development cycle? 2️⃣ - In what situations could you tolerate the system working only 10-20% of the time? For me, the answer to both questions could be: Replacing outsourced development offshore. 🌍 Imagine an organization where precise specs are created in-house and then sent to offshore teams, expected to deliver PRs. This isn’t the typical Tech Native setup, but it exists—and could be a perfect fit for automation given the current model performance. ⚙️ You would send your spec to the agent first, and only if it fails, would you send it offshore—similar to Call Center automation. The human becomes the fallback when the agents can’t deliver. 🔁 This workflow might sound generic, but I believe the quality and level of detail an organization is used to producing in their current dev cycles is critical. In fact, I intuitively think specs are often far more detailed when outsourcing versus working with an internal software team sitting right next to you. 📋✍️ What do you think? 🤔 Do you know any companies specifically focused on automating software development for organizations that rely heavily on outsourcing? PS: Not sure I would trust chatgpt to draw me a flow chart. See below 🤦♂️
Partner Deep tech & Climate
2wInteresting point. As you write, great for companies that are used to creating detailed specs. Interestingly pretty design(er) led companies do it like that, specify everyone and then hand it over for production (AirBnB et all as I’ve understood it, at least for things that are closer the UI). Specifying everything is aoooo tedious though. I’ve worked my fair share with offshoring and found it pretty hard. Only after onboarding team members from the offshore dev company into the org, using the same tools, I’ve gotten it to work. Even if you have relatively strict standards for specifying, a lot gets lost in translation. It takes time to “hand shake” and find a communications style that works on both ends. Taking it one step further could be pretty cool - allow mid-LOD on the specs and then use nature language (on Slack?) or even something video call like with the AI agent to further explain or even collaborate on what your are building.