Claude Code ships with nine built-in tools: Bash, Read, Write, Edit, Glob, Grep, WebFetch, Task, and Plan. For a single agent working on a single codebase, they do the job. But developers building multi-agent systems on top of Claude Code are finding that the defaults fall apart once you move past that basic setup.
The friction points are specific and reproducible.
Cross-Repo Access Is a Manual Chore
Claude Code's file tools are scoped to one project directory. If you want an agent to read code from another repository - say, checking how an open-source library implements something before writing an integration - you have to manually configure permissions each time. There is no "go explore this other codebase and report back" capability out of the box. For teams running agents that need to coordinate across multiple repos, this creates a bottleneck that defeats the purpose of automation.
WebFetch Summarizes When You Need Raw Data
WebFetch does not return raw page content. It runs as a subagent that fetches a URL and hands back a summary. That is fine for quick lookups, but it is a problem when your agent pipeline needs to parse structured data from a page, extract specific code snippets, or process the full HTML. The summarization layer means you lose detail, and there is no built-in way to get the original content instead.
These are not edge cases. They are the exact scenarios that come up the moment you try to build anything resembling a production agent system: agents that research across codebases, agents that scrape and process web data, agents that need to hand off precise context to other agents rather than summaries.
The practical response from developers has been to replace the default tools with custom implementations - writing their own file access tools with broader permissions, swapping WebFetch for direct HTTP clients, and building coordination layers that Claude Code does not provide natively.
This pattern points to a real gap in the current Claude Code architecture. The tool defaults assume a single developer, working in one repo, doing one thing at a time. That matches the original pitch, but the way people actually use Claude Code has already outgrown it. Multi-agent orchestration, cross-project research, and raw data pipelines are not niche use cases anymore - they are where the most ambitious Claude Code users are spending their time.
Anthropic has been iterating on Claude Code quickly, adding hooks, background agents, and MCP support in recent months. Whether the default tool set gets an upgrade to match these multi-agent patterns is worth watching, because right now the gap between "works for one agent" and "works for a fleet" is wider than it should be.