Case Study

Peek-a-Boo

A minimalist AI agent for surgical file search and extraction.

Focused prototype Python CLI Agent Tooling Code Audit Gemini
Peek-a-Boo example mission run

System anatomy

  1. Inputs

    • Target codebase tree
    • Natural-language mission
    • Token + step budgets
    • Allowed tool surface
  2. Core

    • Python agent loop
    • Focused grep / ls primitives
    • Mission-shaped prompts
    • Bounded traversal
  3. Outputs

    • Answers with evidence paths
    • Run telemetry
    • Token-cost summary
    • Reproducible benchmark runs
Constraints
  • Small tool surface by design
  • Hard step + token caps
  • No embeddings
  • No whole-repo context

Why it exists

A lot of code search tasks do not need a full-repo embedding or a giant prompt, they need a cheap agent that can ask narrower questions in sequence. Peek-a-Boo exists for that narrow but common class of work: find the relevant files, inspect just enough context, and produce an answer without turning the entire repository into model input.

Technical center

Peek-a-Boo keeps the tool surface intentionally small, leaning on grep-like operations and focused traversal so the model spends tokens on synthesis instead of raw repository context. The constraint is the product: by limiting available actions and bounding search, the agent has to behave more like a careful code auditor than a broad summarizer.

Current proof points

The project already behaves like a benchmark-shaped prototype: four built-in missions, a generated target codebase, safety limits that prevent token blowups, and a deliberately narrow tool surface that makes the token-saving argument inspectable in code rather than only in prose. That makes it a useful experiment in agent ergonomics because the success criterion is not model cleverness in the abstract, but whether the tool loop finds the right evidence with less context waste.