Agent Development Infrastructure and Integration

A GitHub repository 'awesome-harness-engineering' compiles 12 categories for agent scaffolding. - The list's categories include Agent loop, task decomposition, context compression, tool interface design, permissions, sandbox isolation, validation, observability, and human intervention. - It gathers classic articles from OpenAI and Anthropic on Codex cycles and tool design/permissions. - Reference implementations, tutorials, evaluation methods, and templates are provided. - A quote notes each component exists because the model can't do it, and good design makes them eventually redundant. - Chinese-language documentation is available. - A separate opinion argues services should expose MCP/CLI rather than forcing built-in agents, as users only use 2-3 agents.
Key insights
Curated Agent Scaffolding Resource Highlights Infrastructure Over Model Challenges
The repository identifies 12 categories where scaffolding is needed because the model itself lacks capabilities, such as context compression, tool interface design, permissions, and sandboxing. Classic articles from OpenAI and Anthropic are included to illustrate these design challenges. This indicates that making an agent 'usable' rather than just 'run' requires attention to surrounding infrastructure.
Services Should Expose MCP/CLI for Agent Interoperability Instead of Forcing Built-in Agents
A user argued that people typically use only 2-3 agents, so it's inefficient for each app to embed its own agent. The optimal approach is for services to provide MCP or CLI interfaces, allowing external agents to access the service, following open integration patterns.