Brokers send everything on WhatsApp and leads waited for a human

Agents that can call tools are useful and dangerous in equal measure. The team wanted the useful half: a message becomes a draft listing, a stale listing gets a polite nudge, a buyer gets a first reply within minutes.

What it did not want was an agent that loops, sends twice, or edits something nobody asked it to edit.

• ✦A harness with a fixed step budget, a whitelist of tools per agent, structured logging of every tool call, and a hard stop when the budget runs out.
• ✦WhatsApp capture: turns incoming text, photos and video into a draft listing with the fields it is confident about and asks the broker for the rest.
• ✦Broker follow-up: finds listings that need broker input and drafts the ask; a human approves sends where the stakes are higher.
• ✦Lead replies: drafts the first reply to a buyer enquiry from the listing facts; replies are reviewed by ops before going out.
• ✦An eval harness per agent (chatbot, listing assistant, photo agent, ops agents) with recorded cases that run before changes merge.

• ✦Each agent is a plan over whitelisted tools. The harness runs the loop, records every step, and refuses tool calls outside the whitelist.
• ✦Side effects go through one guarded sender per channel, so a template that is not configured means nothing is sent, and a resend token stops duplicates.
• ✦Evals replay recorded inputs and compare outputs to expected outcomes; a regression blocks the change.