VSCode Team Introduces Five Pillars of Agent-First Development
TL;DR · AI Summary
VSCode team proposes five pillars of Agent-First Development: model selection, action boundaries, context, prompt precision, and tool control, emphasizing the shift from human+editor to human+Agent+editor development paradigm, improving AI programming efficiency through fine-grained configuration.
Key Takeaways
- Copilot provides four levels of thinking depth (Low/Medium/High/Auto) to match d
- Ask→Plan→Agent forms progressive trust ladder to avoid out-of-control risks
- Professional prompts should include goal+scope+constraints+acceptance criteria a
Outline
Jump quickly between sections.
Choose appropriate thinking depth level based on task complexity rather than blindly pursuing large models.
Low level applies to mechanical tasks like completion formatting, high level for architecture design and complex bug handling.
Establish progressive trust ladder through Ask/Plan/Agent three modes to control AI agent permission scope.
Explicitly providing codebase context is key to ensuring AI understands projects and avoiding errors due to information gaps.
Professional prompts should contain goal, scope, constraints, and acceptance criteria for precise output.
Allocate tool permissions as needed, dangerous operations require explicit approval, ensuring observability of tool chains.
Mindmap
See how the topics connect at a glance.
查看大纲文本（无障碍 / 无 JS 友好）
- Agent-First Development五大支柱
- Model模型选择
- 四档思考深度
- 匹配任务需求
- Harness行动边界
- Ask/Plan/Agent模式
- 渐进信任阶梯
- Context上下文
- 自动获取workspace
- 手动指定关键文件
- Prompt提示精度
- 目标+约束+验收标准
- 专业四要素
- Tools工具控制
- 权限分配
- 可观测性
Highlights
Key sentences worth saving and sharing.
Models are not 'bigger is better', but 'matching tasks'.
Copilot provides four levels of thinking depth: Low pursues latency, Medium balances speed and quality, High prefers accuracy over speed.
From Ask→Plan→Agent is the power transfer from 'I lead' to 'it leads'.
Most cases of 'Agent broke the code' root cause is not that the model is bad, but that it didn't see key files before starting to guess.
Professional user prompts typically include: what to do (goal), where to do it (scope), what not to do (constraints), how to know when complete (acceptance).
The subject of writing code is shifting from 'human+editor' to 'human+Agent+editor'.
- Model — Choose the depth of thinking
Models are not "bigger is better," but rather "match the task."
Copilot provides four levels of thinking depth: Low / Medium / High / Auto: · Low: mechanical tasks like completion, renaming, formatting, prioritizing latency · Medium: regular feature development, balancing speed and quality · High: architecture design, complex https://t.co/VnhJWBXBwh" / X
VSCode team introduces the five pillars of Agent-First Development 1. Model — Choose the depth of thinking Models are not "bigger is better," but rather "match the task." Copilot provides four levels of thinking depth: Low / Medium / High / Auto: · Low: mechanical tasks like completion, renaming, formatting, prioritizing latency · Medium: regular feature development, balancing speed and quality · High: architecture design, complex bugs, cross-file refactoring, better slow than wrong · Auto: let the system automatically schedule based on task complexity The real insight: thinking depth has costs (time, tokens, user patience). Mature Agent-First users will actively select "gears" for tasks instead of defaulting to High mode for everything. 2. Harness — Agent's action boundaries Harness refers to what an Agent can and cannot do within its runtime environment. The Copilot Chat panel is a harness that determines whether the Agent merely "talks" or actually "takes action." Three modes form a progressive trust ladder: · Ask: conversation only, no file changes - exploration, asking questions, learning code · Plan: list steps, await your approval - solution discussion, high-risk changes · Agent: autonomous planning, execution, iteration - clear direction, implementation-needed tasks Key insight: moving from Ask → Plan → Agent represents a power transfer from "I lead" to "it leads." Forcing Agent mode for ambiguous tasks is like letting someone unfamiliar with the situation make decisions for you—loss of control is inevitable. 3. Context — Agent's "field of view" The model itself doesn't know your codebase. Context is the prerequisite for it to do things correctly. Copilot acquires context in two ways: · Automatic: search workspace, read relevant files · Manual (more reliable): explicitly specify using # codebase, # file, # fetch This is the most underestimated element. Most cases of "Agent broke the code" stem not from model inadequacy, but from the agent starting to guess without seeing key files. Explicitly providing context isn't redundant—it's controlling variables. 4. Prompt — Precision of intent Prompt = goal + constraints + acceptance criteria. "Help me add login functionality" versus "implement email login in the auth/ directory using existing SessionService, reuse LoginForm component, show errors via toast, don't introduce new dependencies" yield completely different results. Professional user prompts typically include: · What to do (goal) · Where to do it (scope) · What not to do (constraints) · How to know when complete (acceptance) Fuzzy input can only yield generic output. 5. Tools — Agent's "hands" Tools are the actual capabilities an Agent can invoke: reading files, running commands, searching, calling APIs, running tests… More tools mean greater power, but also larger blast radius. Agent-First discipline requires: · Give it needed tools, not unnecessary ones · Dangerous operations (delete database, push to main, modify prod config) should require explicit approval · Toolchains should be observable—what commands did the Agent run, what files did it change, must be visible and reversible Overall framework: five dials, not five steps Poor results = f(model gear mismatch × harness mode error × missing context × vague prompt × insufficient/excessive tools) The essence of Agent-First Development acknowledges one thing: the primary actor in coding is shifting from "person + editor" to "person + Agent + editor." Human roles transition from "typist" to "model selector, context provider, intent definer, boundary setter." VS Code Learn series videos youtube.com/playlist?list=
Quote
Microsoft Developer
@msdev
9h
The @code team released a new Introduction to Agent-First Development series. It breaks down these 5 pillars behind great agent results 