Shell + Skills + Compaction: Tips for long-running agents that do real work
Practical patterns for building with skills, hosted shell, and server-side compaction in the Responses API.
AI Agent Optimization
Summary
Amp enables lazy loading of MCP tools via Agent Skills, injecting tool definitions into the context window only upon invocation. This reduces token usage from 17k to 1.5k by avoiding upfront loading of unused tools.
Highlights
- Tools load into context only when the associated skill is invoked.
- Configuration uses an mcp.json file to specify included tools.
- Reduces token usage significantly by excluding unused tool definitions.
- Prevents context window overload from large MCP server toolsets.
auto-generated
Context
Audience
AI developers and engineers using large language models
DomainSoftware Engineering
Formattechnical documentation
Accessfree online
Topics
Related
AmpClaude CodeMCP ToolsAgent Skillschrome-devtools-mcp
