ManyChat AI Agent

Appearance

ADR-0007 — generateObject rather than a tool loop for v1 ​

Status: superseded by ADR-0010 · Date: 2026-09-14

Context ​

The AI SDK offers ToolLoopAgent, which runs the full call/execute/repeat cycle. It is the obvious thing to reach for in something called "an agent". The v1 scope is answer-from-catalog and escalate: there are no tools to call.

Decision ​

Implement the AgentRunner port with a single generateObject call returning the AgentReply schema. No tool loop in v1.

Consequences ​

  • One model call instead of a loop of them, which matters against an 8s budget — a tool loop's step count is unbounded in latency terms.
  • The escalation signal is a validated field rather than a string parsed out of prose, so routing to a human is mechanical.
  • Adding tools later means a second AgentRunner implementation behind the same port; no caller changes.
  • Gives up built-in multi-step reasoning. If the agent later needs to check live availability or create a booking, this decision gets revisited — that is the trigger, not a general desire for the agent to be more agentic.