New York Code Summit: Insights on Agents and Engineering Practices

The New York Code Summit focused on key points regarding agent engineering practices, organizational change, and governance.
Key insights
Organizational Structure and Development Processes Need Reconstruction to Realize AI Value
A McKinsey report indicates that most companies embed AI into existing agile processes, resulting in approximately a 10% increase in revenue. The conference suggested shifting from 'two-pizza teams' to smaller 'single pizza teams', abandoning user story-centric processes in favor of specification-driven development; without changing the organizational structure, true AI adoption cannot be achieved.
Codebase Quality and Incentive Design Determine the Success or Failure of AI Engineering
Related analysis from Stanford points out that simply issuing tokens to developers is unfeasible, as there exists a 'valley of death' between high usage rates and performance decline. A chaotic codebase can accelerate entropy through AI, increasing rework by approximately 2.5 times. Perspectives from companies like Capital One emphasize the need to stop using obscure technologies; if agents cannot understand legacy code, neither can humans.
IDE and Coding Methods Face Transformation, with Clear Trends Towards Productionization
Several speakers predict that traditional IDEs will be replaced, stating that IDEs will decline by 2026. Large context windows are not a panacea, and 'vibe coding' is being pushed into production environments. Some pointed out that if code is still being handwritten next year, it would be akin to using tools from the era of hand drills.
Building 'Protection Layers' and Middle Platform Abstractions to Avoid Rebuilding Runtime Environments
The released alpha version demonstrated that models are not a moat but require a 'protection layer' abstraction: constructing an SDK that allows agents to use terminals and computing resources to avoid rebuilding the runtime environment with each model update. The conference also suggested building a middle platform to prevent multiple teams from redundantly implementing the same PR bots.
Context Engineering and Token Budgeting Can Significantly Enhance Agent Performance
Participants shared practices: allowing models to use token budgets to 'think' and employing context engineering to enable agents to eliminate their own garbage outputs, with reports indicating a performance increase of approximately 39%. They advocated stopping the blind filling of context windows and adopting garbage collection-style context management.
Multi-Agent Collaboration, Testing, and Governance Are Key to Implementation
Demonstrated multi-agent collaboration tools and CLI (such as CodeMachine CLI) can convert requirement documents into production-grade code while handling architecture, writing, testing, and deployment in parallel. Participants discussed the necessity of generating Playwright tests with subagents to monitor the main loop, avoiding superficial engineering, and noted that prompt injection and production environment security remain unresolved issues.
Source posts
- Original post by @Yangyixxxx on X
- Original post by @Yangyixxxx on X
- Original post by @Yangyixxxx on X
- Original post by @Yangyixxxx on X
- Original post by @Yangyixxxx on X
- Original post by @Yangyixxxx on X
- Original post by @GitHub_Daily on X
- Original post by @Yangyixxxx on X
- Original post by @Yangyixxxx on X
- Original post by @Yangyixxxx on X
- Original post by @Yangyixxxx on X
- Original post by @Yangyixxxx on X
- Original post by @Yangyixxxx on X
- Original post by @Yangyixxxx on X
- Original post by @Yangyixxxx on X