How TaxPilot works

What is an agent here?

The agent follows a reason → act → observe loop. It reads your question, picks one of 43 tax calculators, looks at the result, and repeats until it can answer. It never invents numbers; every figure comes from the engine.

Why MCP?

Model Context Protocol is the same interface the upstream engine already exposes. We connect to it in-memory, so there is no separate server to run and no network hop. If we ever need to swap to a remote engine, only the transport changes.

Why numbers never come from the AI

The AI only sees tool names, descriptions and your question. It chooses tools and fills in arguments. The engine runs the math and returns markdown tables. The AI then explains those results in plain English.

Guardrails

• Maximum 8 agent steps and 6 real tool calls per run.
• Input capped at 1,500 characters; only 10 recent messages kept.
• In-memory per-IP rate limit (best-effort on serverless).
• Prompt caching to keep costs low across repeated questions.
• Recorded replay fallback when the live agent is unavailable.

What I'd build next

• A multi-step planner that breaks complex questions into sub-tasks.
• A verifier agent that double-checks tool outputs for consistency.
• Clarifying questions when the user omits key facts.
• An evaluation suite that scores answer accuracy on known examples.