Direct Tool Selection

Direct Tool Selection Protocol (Zero-Exploratory Probing)

Overview

When asked to locate or modify a specific piece of code ("Fix the JWT expiration check in the auth service"), unoptimized agents perform a multi-step Exploratory Probing Sequence:

1. Turn 1: Call list_dir("/") to explore folders.
2. Turn 2: Call run_command("find src/ -name *auth*") to locate the file.
3. Turn 3: Call view_file on the whole file.
4. Turn 4: Execute the edit.

This 4-turn sequence wastes 3 full API roundtrips, incurs 15 to 20 seconds of latency, and burns 6,000+ context tokens before writing a single line of code.

The Direct Tool Selection Protocol maps user intents directly to the single highest-specificity tool on Turn 1 (e.g., executing grep_search(Query: "jwt.verify") immediately).

Exploratory Probing Loop vs. Direct Tool Invocation

┌─────────────────────────────────────────────────────────────┐
│                 Tool Selection Trajectory                   │
│                                                             │
│  Exploratory Probing Loop (4 Turns / 6,500 Tokens):         │
│  • Turn 1: `list_dir` ──► Dumps directory tree (800 tokens) │
│  • Turn 2: `run_command(find)` ──► Shell output (200 tokens)│
│  • Turn 3: `view_file` ──► Dumps entire file (3,500 tokens) │
│  • Turn 4: `replace_file` ──► Performs edit                 │
│  ↳ 4 Turns, 22 Seconds Latency                              │
│                                                             │
│  Direct Tool Selection Protocol (1 Turn / 280 Tokens):      │
│  • Turn 1: `grep_search("jwt.verify")`                      │
│    ↳ Pinpoints `src/auth.ts:42` with 1 exact tool call!     │
│  • Turn 2: `replace_file_content`                           │
│  ↳ 2 Turns Total, 2.5 Seconds Latency (8.8x Faster!)        │
└─────────────────────────────────────────────────────────────┘

The Direct Tool Selection Decision Matrix

Always map operational goals directly to their optimal specialized tool:

The 3 Rules of Turn-1 Tool Execution

1. Grep First, Never Browse

If you are searching for where a variable, error code, or API route is defined, invoke grep_search on Turn 1. Never explore the directory tree folder-by-folder.

2. Bypass Browser Subagent for Static URLs

If you need documentation, blog posts, or API reference text from a public URL, call read_url_content. Use browser_subagent only if the site requires interactive clicking, form submissions, or JavaScript authentication.

3. Combine Line Lookup with Bounded Inspection

When grep_search returns match at line 142, call view_file(StartLine: 135, EndLine: 160). Never omit the start/end lines.

Benchmark Comparison

Executing 40 codebase search and bug-fix operations:

Agent Operational Directive

MANDATORY: Agents must select the single most specific tool on Turn 1. Never execute exploratory directory listing or shell searching if grep_search or line-bounded view_file can target the exact symbol directly.