Agent-Laufzeiten und MCP recherchieren

AI Agent, Runtime & MCP Ecosystem Research

AI and agent repositories are especially prone to "prompt-demo hype" — projects that look magical in a 30-second video but consist of brittle regexes, unbounded token loops, or hardcoded happy paths. Always dissect the runtime mechanics underneath.

Workflow

1. Dissect the agent execution loop and state model:
   • Locate the actual turn/step loop in source code and verify:
     • State persistence & resumability: Is run state checkpointed durably between tool calls, or held only in volatile in-memory variables?
     • Termination & loop guards: How are max turns, token budgets, repeated identical tool errors, and AbortSignal cancellations enforced?
     • Structured output & tool parsing: Does it use native provider tool-use / structured-output schemas with validation and repair, or fragile string splitting?
2. Evaluate Model Context Protocol (MCP) servers and integrations:
   • Inspect MCP implementations across protocol and security dimensions:
     • Transport support: stdio subprocess lifecycle (clean shutdown, stderr logging vs. stdout JSON-RPC frame corruption), SSE, or Streamable HTTP.
     • Tool & resource schema quality: Are input JSON Schemas strict, descriptive, and bounded in output size, or do they dump multi-megabyte raw payloads into the model context?
     • Permission & secret boundaries: How are credentials passed? Does the MCP server expose unrestricted filesystem, shell, or network access without path/command allowlists?
3. Analyze memory, RAG, and context-management architectures:
   • Trace how context is assembled on each turn:
     • Prompt cache friendliness: Does the system preserve a stable prefix (system prompt + tool definitions + frozen history) so Anthropic/OpenAI/Gemini prompt caching works, or does it inject dynamic timestamps/UUIDs at the top of every prompt?
     • Compaction & retrieval: How are long tool outputs truncated, summarized, or paged? What storage engine and embedding/reranking dependencies are required?
4. Extract reusable patterns over bloated wrappers:
   • Many "agent frameworks" wrap 50 lines of provider SDK calls in 15 layers of abstract classes. When a project's core value is a clever prompt structure, compaction heuristic, or tool schema, recommend extracting the pattern or adapting the minimal logic rather than adopting a rigid framework that fights our orchestrator.