agent-tool-design（Agent 工具设计）

Agent Tool Design

The Agent Tool Contract — 5 principles for designing tools that agents call reliably.

The 5 Principles

Principle 1: Predictable Signature

Tools must have typed, named parameters with clear required/optional distinction. No positional ambiguity.

Good:

// Clear, named, typed
function searchCode({ query, limit = 20, type = 'semantic' }) { ... }

Bad:

// Positional, ambiguous
function searchCode(q, n, t) { ... }

Principle 2: Rich Errors

Errors must include: error code (machine-readable), message (human-readable), context (debugging data).

Good:

throw {
  code: 'FILE_NOT_FOUND',
  message: `File not found: ${path}`,
  context: { path, cwd: process.cwd() },
};

Bad:

throw new Error('not found'); // No context for agent to act on

Principle 3: Token-Efficient Output

Tools return structured minimal data. No prose explanations, no redundant wrapping, no verbose status messages. Agents format output themselves.

Good:

return { files: ['a.js', 'b.js'], total: 2 };

Bad:

return { status: 'success', message: 'Found 2 files successfully', data: { files: [...], metadata: {...} } };

Rule of thumb: If the output contains prose an agent would re-read to extract facts, it's too verbose.

Principle 4: Idempotency

Tools must be safe to retry. Running a tool twice should produce the same result as running it once.

Good:

// Upsert instead of insert
db.upsert({ id, ...data });
// mkdir -p instead of mkdir
fs.mkdirSync(path, { recursive: true });

Bad:

// Fails on retry
db.insert({ id, ...data }); // duplicate key error
fs.mkdirSync(path); // EEXIST error

Principle 5: Graceful Degradation

Partial success > hard failure. Return what succeeded with a clear indication of what didn't.

Good:

return {
  succeeded: ['file1.js', 'file2.js'],
  failed: [{ file: 'file3.js', reason: 'PERMISSION_DENIED' }],
  partial: true,
};

Bad:

// One file fails -> entire batch throws
throw new Error('Failed to process file3.js');

Anti-Pattern Table

Review Checklist

Before shipping any tool:

[ ] Parameters are named (not positional)
[ ] Required vs optional params are explicit
[ ] All error paths return { code, message, context }
[ ] Output contains no prose — only structured data
[ ] Tool is idempotent (safe to retry)
[ ] Partial failure returns partial results, not throws
[ ] Return shape is documented in JSDoc or TypeScript types
[ ] Token budget for output estimated (< 500 tokens for standard tools)

Iron Laws

1. ALWAYS use named parameters — never positional arguments in tool signatures; positional args break on refactor and create ambiguity for agents.
2. ALWAYS include machine-readable error codes — never surface plain string errors only; agents need {code, message, context} to handle errors programmatically.
3. NEVER mix reads and writes in the same tool — read tools that trigger side effects confuse agents and prevent safe retries.
4. ALWAYS design for idempotency — retry must produce the same result as the first call; use upsert semantics and mkdir -p patterns.
5. ALWAYS return partial results on partial failure — never let one failing item abort the entire batch; return {succeeded, failed, partial: true}.

Integration

• Used by: tool-creator skill when designing new tools
• Reviewed by: code-reviewer agent during tool PRs
• Pairs with: dynamic-api-integration skill (consuming external tools)
• Complements: agent-evaluation skill (evaluating tool output quality)

Memory Protocol (MANDATORY)

Before starting: Read .claude/context/memory/learnings.md

After completing:

• New pattern -> .claude/context/memory/learnings.md
• Issue found -> .claude/context/memory/issues.md
• Decision made -> .claude/context/memory/decisions.md

ASSUME INTERRUPTION: If it's not in memory, it didn't happen.