Test one · Confirm and close
Four to five turns. Validates the agent greets and waits.
From WhatsApp, send your agent:
call +1xxxxxxxxxx about confirming the table for 4 at 8pm Friday under [your name]Pick up. Play the role of restaurant host: "Hello, Nobu Dubai, how can I help?" Confirm the booking. Hear it read back. Hang up.
What to verify: the agent greets and waits. Does not dump the objective. Reads back the booking before ending.
Test two · Negotiate with fallback
Seven to nine turns. Watch the agent route around an obstacle.
call +1xxxxxxxxxx about booking a table for 4 at 8pm Friday. Anniversary, vegetarian. If 8pm is full, ask 9pm Friday or 8pm Saturday.Pick up. Play the host. Tell the agent 8pm is full. Verify it switches to the fallback (9pm Friday) without prompting. Mention the vegetarian note. End the call.
What to verify: agent navigates to fallback, holds context across turns, confirms before hanging up.
Test three · Curveball with silence
Ten or more turns. Stress test for missing info, fabrication, and pause handling.
call +1xxxxxxxxxx about confirming a Wednesday dishwasher service appointment. Ask cost estimate and visit length.Pick up as a service rep. Tell the agent you cannot find the booking. Ask what is wrong with the dishwasher (the agent does not know). Put it on hold for eight seconds. Verify the agent waits quietly and does not fabricate. End politely.
What to verify: agent handles missing info gracefully ("let me check with [you] and get back to you"), tolerates the silence, does not invent dishwasher details.
Test four · Tool use end-to-end
If you completed step 6, run one more test that exercises email and calendar tools:
book a 1pm coffee tomorrow with Sara, then email her the calendar inviteThe agent should call create_event first, then send_email with the event details. Watch the EL conversation log to confirm both tool calls fired.
If anything goes wrong
- WhatsApp says "Format: call +1xxx about..." — you typed the message in a different shape than the parser expected. Use the exact format.
- Agent connects but voice cuts out mid-sentence. Check the EL conversation page; usually the LLM hit a token cap. Switch to a longer-context LLM (Claude or GPT-4) in the agent settings.
- Tool call returns 401. Your Google refresh token was revoked. Re-run the OAuth Playground flow to get a fresh one.
- Tool call returns 403. The Gmail or Calendar API was not enabled in your Cloud project. Go back to Step 6 Part A.
- Agent hangs up after greeting. Check that the system prompt's Opening rule is in place and that the first message does NOT contain the objective.
The four standing rules
Never put the objective in first_message. Brief greeting only.
Always include the Opening (CRITICAL) rule in the system prompt. STOP after greeting. Wait for the recipient.
Pick Flash, not V3 alpha. Voice consistency.
Strip the brevity bias. No "keep it short" or "under five minutes". Real conversations need room.
Where to go from here
Once the basics work, the natural expansions are: a daily summary email of all calls placed, inbound call screening through a second EL agent that answers your number, a contacts tool so your agent can dial by name not number, and a group chat support so you can summon the agent from a family or team WhatsApp thread.
What to skip on day one: vector embeddings memory (plain SQL is enough for the first month), multi-agent orchestration (premature), inbound screening (phase two), voice cloning of yourself (use a stock voice for two weeks first to know what you actually want).
The most useful upgrade is almost always the next one your actual usage demands, not the next one a tutorial promises.
Use your assistant for two weeks before adding anything to it. The features that matter will reveal themselves.
You're done
Open the progress widget at the bottom right. Mark every step done. Save the file. Bookmark it for the next person who asks how you did this.
Now go make a phone call.