LinkedIn·Thursday, 20 August 2026·6d ago

🧭 𝗙𝗿𝗼𝗺 𝘃𝗶𝗯𝗲 𝗰𝗼𝗱𝗶𝗻𝗴 → 𝗱𝗼𝗰𝘂𝗺𝗲𝗻𝘁-𝗳𝗶𝗿𝘀𝘁 → 𝘀𝗽𝗲𝗰-𝗱𝗿𝗶𝘃𝗲𝗻 𝗰𝗼𝗱𝗶𝗻𝗴 I’ve been using AI coding agents…

Eric Charles

CEO / Founder at Datalayer ≡ Managed AI Agents for Data Analysis ✨🤖 #faster #cheaper #collaborative

🧭 𝗙𝗿𝗼𝗺 𝘃𝗶𝗯𝗲 𝗰𝗼𝗱𝗶𝗻𝗴 → 𝗱𝗼𝗰𝘂𝗺𝗲𝗻𝘁-𝗳𝗶𝗿𝘀𝘁 → 𝘀𝗽𝗲𝗰-𝗱𝗿𝗶𝘃𝗲𝗻 𝗰𝗼𝗱𝗶𝗻𝗴 I’ve been using AI coding agents extensively for more than a year. For a long time, my workflow was mostly 𝘷𝘪𝘣𝘦 𝘤𝘰𝘥𝘪𝘯𝘨: define a reasonably well-scoped feature, sometimes write a document about it, give the context to the agent, and iterate. And it worked brilliantly. ⚡ For me, the productivity boost has easily been 𝟭𝟬𝘅. But I’m now trying to make that workflow much more structured. 📄 𝗕𝗲𝗳𝗼𝗿𝗲 𝘄𝗿𝗶𝘁𝗶𝗻𝗴 𝗰𝗼𝗱𝗲, 𝘄𝗿𝗶𝘁𝗲 𝘁𝗵𝗲 𝗱𝗼𝗰𝘂𝗺𝗲𝗻𝘁. A recent example is our new Code Sandboxes provider work at Datalayer. Instead of immediately asking the coding agent to implement the feature, I first asked it to: 🔍 review the documentation ✍️ challenge and improve it 🧩 identify what was underspecified 📐 clarify the expected behavior ⚙️ and only then start implementing The result is now live. 🚀 𝗗𝗮𝘁𝗮𝗹𝗮𝘆𝗲𝗿 𝗖𝗼𝗱𝗲 𝗦𝗮𝗻𝗱𝗯𝗼𝘅𝗲𝘀 𝗻𝗼𝘄 𝘀𝘂𝗽𝗽𝗼𝗿𝘁 𝗯𝗼𝘁𝗵 𝗖𝗣𝗨 𝗮𝗻𝗱 𝗚𝗣𝗨 𝗲𝗻𝘃𝗶𝗿𝗼𝗻𝗺𝗲𝗻𝘁𝘀 𝗽𝗿𝗼𝘃𝗶𝗱𝗲𝗱 𝗯𝘆 𝗠𝗼𝗱𝗮𝗹 𝗮𝗻𝗱 𝗞𝗮𝗴𝗴𝗹𝗲. That means developers and agents can work with different remote compute environments depending on the workload — from CPU-based tasks to GPU-intensive AI workloads. ⚡🤖 And the document that helped drive the implementation is now part of the actual Datalayer documentation: 📚 https://lnkd.in/eQ_enmqf That is the part I find particularly interesting. The document wasn’t a disposable prompt. It became: 📄 the specification for the agent 🧠 a way to clarify the feature before implementation 📚 and part of the documentation shipped with the product My workflow is increasingly moving toward: 💭 𝗩𝗶𝗯𝗲 𝗰𝗼𝗱𝗶𝗻𝗴 ↓ 📄 𝗗𝗼𝗰𝘂𝗺𝗲𝗻𝘁-𝗱𝗿𝗶𝘃𝗲𝗻 𝗰𝗼𝗱𝗶𝗻𝗴 ↓ 📐 𝗦𝗽𝗲𝗰-𝗱𝗿𝗶𝘃𝗲𝗻 𝗱𝗲𝘃𝗲𝗹𝗼𝗽𝗺𝗲𝗻𝘁 I now want to push this further and explore tools such as GitHub’s 𝗦𝗽𝗲𝗰 𝗞𝗶𝘁: 🔗 https://lnkd.in/eTs6c45w Including its extension model: 🧩 https://lnkd.in/e9kn226J I’ve also opened a discussion in 𝗱𝗮𝘁𝗮𝗹𝗮𝘆𝗲𝗿/𝗮𝗴𝗲𝗻𝘁-𝗿𝘂𝗻𝘁𝗶𝗺𝗲𝘀 to engage with the community around where we could take this: 💬 https://lnkd.in/eaNGTYdK I’d love to hear how others are approaching spec-driven development with coding agents. The more capable agents become, the more leverage seems to move from 𝘸𝘳𝘪𝘵𝘪𝘯𝘨 𝘵𝘩𝘦 𝘤𝘰𝘥𝘦 toward 𝗱𝗲𝗳𝗶𝗻𝗶𝗻𝗴 𝗽𝗿𝗲𝗰𝗶𝘀𝗲𝗹𝘆 𝘄𝗵𝗮𝘁 𝘀𝗵𝗼𝘂𝗹𝗱 𝗯𝗲 𝗯𝘂𝗶𝗹𝘁. 🧪🤖 #AICoding #CodingAgents #SpecDrivenDevelopment #Jupyter #GPU #AIEngineering #Datalayer 📐 Create a `spec-kit` plugin · Issue #116 · datalayer/agent-runtimes https://github.com/datalayer/agent-runtimes/issues/116