Rebeca Moen
Apr 10, 2026 19:10
Anthropic engineers detail how they build and refine AI agent tools for Claude Code, introducing progressive disclosure techniques that shape AI development.
Anthropic has pulled back the curtain on how its engineering team designs tools for Claude Code, the company’s AI-powered software development assistant. The detailed technical breakdown, published April 10, offers rare insight into the iterative process behind building effective AI agent systems.
The $380 billion AI safety company’s approach centers on what engineer Thariq Shihipar calls “seeing like an agent” — essentially understanding how an AI model perceives and interacts with the tools it’s given.
Trial and Error with AskUserQuestion
Building Claude’s question-asking capability took three attempts. The team first tried adding a question parameter to an existing tool, which confused the model when user answers conflicted with generated plans. A second attempt using modified markdown formatting proved unreliable — Claude would “append extra sentences, drop options, or abandon the structure altogether.”
The winning solution: a dedicated AskUserQuestion tool that triggers a modal interface, blocking the agent’s loop until users respond. The structured approach worked because, as Shihipar notes, “even the best designed tool doesn’t work if Claude doesn’t understand how to call it.”
When Tools Become Constraints
The team’s experience with task management reveals how model improvements can render existing tools obsolete. Early versions of Claude Code used a TodoWrite tool with system reminders every five turns to keep the model on track.
As models improved, this became counterproductive. Claude started treating the todo list as immutable rather than adapting when circumstances changed. The solution was replacing TodoWrite with a more flexible Task tool that supports dependencies and cross-subagent communication.
From RAG to Self-Directed Search
Perhaps the most significant shift involved how Claude finds context. The initial release used retrieval-augmented generation (RAG), pre-indexing codebases and feeding relevant snippets to Claude. While fast, this approach was fragile and meant Claude was “given this context instead of finding the context itself.”
Giving Claude a Grep tool changed the dynamic entirely. Combined with Agent Skills — which allow recursive file discovery — the model went from being unable to build its own context to performing “nested search across several layers of files to find the exact context it needed.”
The 20-Tool Ceiling
Claude Code currently operates with roughly 20 tools, and Anthropic maintains a high bar for additions. Each new tool represents another decision point for the model to evaluate.
When users needed Claude to answer questions about Claude Code itself, the team avoided adding another tool. Instead, they built a specialized subagent that searches documentation in its own context and returns only the answer, keeping the main agent’s context clean.
This “progressive disclosure” approach — letting agents incrementally discover relevant information — has become central to Anthropic’s design philosophy. It echoes the company’s broader focus on creating AI systems that are helpful without becoming unwieldy or unpredictable.
For developers building their own agent systems, the takeaway is clear: tool design requires constant iteration as model capabilities evolve. What helps an AI today might constrain it tomorrow.
Image source: Shutterstock
