The Gmail API doesn’t have a “reply to email” endpoint. Let that sink in.

When a user says “reply to Alex’s email,” they’re expressing clear intent. The LLM understands it perfectly. But the API underneath? It has no idea what a “reply” even is.

So your agent has to:
- Figure out which Alex
- Search through messages (40 parameters deep)
- Unpack MIME by hand
- Maintain threading
- Repack everything
- Finally send

This isn’t a prompt-engineering problem.
It’s not a model problem.
It’s an architecture problem.

You’re asking your LLM to translate human intent into low-level API primitives.

Every translation step adds:
- More tokens ($$$)
- More latency (seconds add up)
- More hallucination risk

Teams keep trying to fix this by dumping context into prompts —
“Here’s how MIME works. Here’s the threading format. Here’s…”

Wrong approach. Your LLM is doing plumbing work when it should be doing reasoning work.

The right approach: Build tools that match agent intentions.
One “reply to email” tool. Handles all the complexity internally.

The agent makes one decision, not six.

That’s why 70% of agent projects fail — they’re building API wrappers when they should be building intention engines.

At Arcade.dev, we built that layer — tools that map human intent directly to secure, real-world action.
▿ Show more