🧠 The conversation around AI for developers usually starts with the model. But inside @code, what really shapes the experience is the coding harness: the layer responsible for context, tool calling, agent loops, terminal execution, memory, and more. In this new post, the engineering team dives into how GitHub Copilot in VS Code works behind the scenes. code.visualstudio.com/blogs/2026/05/…

@code Exactly right - it's all about the harness. And each model is trained to use their own harness the best. Keep the agents in their comfort zone. github.com/mcknly/robot-d…

@code This is the right framing. The model is only one layer. The harness decides context, tool use, terminal execution, memory, and review loops. For agents, product quality increasingly lives in the system around the model.

@code this is the part people skip when they say “just use a better model”

now means now. in the grey is now...now playing in theaters everywhere. starring jake gyllenhaal, henry cavill and eiza gonzalez. written & directed by guy ritchie. get your tickets...now! 🎟️

@code models are flashy, but the harness is where the real product decisions live

The model gets all the attention, but the coding layer is where the risk turns into action. Context, tool calls, terminal access, MCP servers, memory, network traffic. That's the surface that needs inspection and evidence, not a nicer chat box. I wrote up the MCP wrapping path for VS Code here: pipelab.org/learn/vscode-i…

@code the harness is where model capability becomes actual work instead of a leaderboard screenshot

The model/harness distinction is underrated. Everyone debates which model is best, but the harness — context management, tool routing, memory scope — is what actually determines if the agent stays on track or goes off the rails. Would love to see more teams publish their harness design decisions.

@code The harness framing is right. The model is almost a commodity now — what differentiates coding AI is the harness: does it know your codebase, handle context degradation, recover from dead ends? Quality of context management is what actually separates good tools from great ones.

@code Harness insight is right but incomplete. What's missing is constraint architecture. Tool calling without permission boundaries is giving root access to a junior dev. Context, tools, memory are the engine. Constraints are the steering wheel.

@code This is ridiculous. You keep messing with my subscription and slapping rate limits while I still have available requests. Went from monthly Opus 4.6 with full access until quota end to restrictions everywhere + 3+ days waiting for reset. Stop screwing over your customers

@code 把AI Agent当同事的前提是：它能独立担责。现在99%的Agent只是在执行指令，不是在做决策。

@code A strong coding harness can make an average model feel powerful, while a weak harness can make a great model frustrating to use. Infrastructure around the model is becoming the differentiator.

@code i hope u guys can improve it a lot, maybe someday will be on pair of tools like cursor, opencode, windsurf.