Claude Advanced Tool Use

Shifting Agent Complexity & Hype Cycles

• Commenters see a recurring cycle: heavy scaffolding (LangChain-style agents) → simpler loops → MCP with large schemas → back to bash/filesystem → now richer tool systems again.
• Some express fatigue at constant reinvention and hype-driven adoption, with good OSS ideas often ignored until a major vendor rebrands them.
• Others note cost and latency will be strong incentives to simplify once capabilities stabilize.

Programmatic Tool Calling & Code as the “Tool Language”

• Strong support for “write code to call tools instead of calling them directly”: lets agents batch calls, pass data between tools, and avoid copying large payloads into the model’s context.
• Several people mention similar prior work (e.g., smolagents, MCP proxies that turn tools into TypeScript/Python APIs executed in a sandbox).
• There’s a broader push toward giving models a real programming environment (Python, bash, TypeScript, Prolog DSLs) instead of verbose JSON schemas, with the model orchestrating tools via code.

Context Management, Tool Search & “RAG for Tools”

• Many criticize loading all tool JSON schemas into context as wasteful and a known anti-pattern. Patterns discussed: tool search tools, skills folders, sub-agents, plans, and delayed “install” of tools.
• Anthropic’s Tool Search is seen by some as “RAG for tools”: offload discovery, then only load a small subset into context. Others argue good architecture (per-state tool sets, sub-agents) already solves this.
• Concern that debugging opaque tool selection will be painful when the wrong tool is silently chosen.

Security, Ecosystem & Optimization Games

• Letting LLMs discover tools on GitHub and execute them is called a security nightmare: sandboxing protects machines but not data exfiltration; curation is urged.
• People anticipate “Tool Engine Optimization” (TEO), promoted tools, and ranking systems analogous to SEO/PageRank.

Alternatives: GraphQL, SPARQL, CLI & Shell-First

• Multiple commenters advocate GraphQL (or SPARQL) as a single, typed, introspectable “super-tool” that avoids dozens of separate MCP tools and supports selective data loading.
• Others emphasize good old CLI tools with --help as simpler, composable interfaces that LLMs can already use, often via a shell tool, sometimes preferred over buggy MCP servers.

Reliability & MCP/Skills Critiques

• Reports that tools, MCP, and skills integrations are brittle: skills often not invoked unless explicitly named; CLAUDE.md and similar configs get ignored due to context rot.
• Some argue MCP is conceptually interesting but currently too buggy for production, pushing teams to build custom protocols or lighter abstractions.