Naveo

STEP 14 / 27

A7 A/B

MCQ · NO COST

Forge is building an MCP for a shift assistant. The PM sent this request: "add a tool that converts informal shift names ('morning', 'night', 'late hours') into the system's formal shift_type ('day', 'night', 'late_night')". What does Forge do?

Why?. optional

Look for: closed contract, explicit fallback, scaffold at the end.

GUEST MODE

You're viewing this lesson as a guest. To save your progress, earn XP, and keep your streak, sign in when you're ready to check.

Costs 1 heart

Tooling everything is an anti-pattern

It's tempting to solve every "logic" with a tool. You get a system that's "modular", "auditable", "testable". But each extra tool:

Adds work for the model. More options to scan. More decisions of "do I use this or that?".
Adds work for you. Another handler to maintain, another schema to version, another piece of the system that can crash.
Slows the agent down. Every tool call is another round (model → tool → model). What the LLM could do in one turn now takes three.

And many times, the "logic" you're encapsulating is something the LLM does for free and better.

The quick test: what does this need?

Ask this question before creating any tool:

Does the operation need access to data, systems, or guarantees the model doesn't have?

Needs	Tool yes	Tool no
Read/write your DB	✓
Call an external API	✓
Side effects (mails, payments, notifs)	✓
Absolute determinism (financial calc)	✓
Real-time data (prices, stocks)	✓
Translate / normalize / rephrase		✓ (system prompt)
Summarize content in context		✓ (in-prompt)
Decide tone / response style		✓ (system prompt)
Simple calculation with visible numbers		✓ (in-prompt)
Classify text into known categories		✓ (in-prompt or output schema)

Real case: Forge almost fell

The ship had an MCP with 23 tools. Forge audited and found 8 of them were "normalizers": normalize_alias, format_date, parse_priority, expand_abbreviation. All of them could be replaced by two paragraphs in the system prompt stating the normalization rules.

Result of deleting the 8: the agent got 1.4x faster, the catalog dropped to 15 tools, and the precision on picking the remaining tools went up (less noise, better decision).

Before writing a tool, ask: does this need the real world, or does it only need the model to think a bit more in the prompt? If it's the second, it's not a tool.