5 steps to build a support agent you can trust
JP Valery
Customer Success Engineer, Resend
Susana de Sousa
Community, Plain
Last Updated
Published On
Scope
01.
Start with one type of request
Pick a request that comes in often, gets the same answer every time, and needs data you can fetch easily. For Resend, that was domain verification.
Tip: Ask engineering for help early. Getting from prototype to production was Resend’s biggest delay.
Architecture
02.
Split the agent into three jobs
Resend calls each of these flows a recipe and runs more than 20 of them.
Understand
Classify the request. If the agent isn’t sure, a person takes it.
Gather
Fetch only the account data this request needs.
Reply
Write the reply from the message, the request type, the data, and your instructions.
Testing
03.
Test before customers see it
Resend trusts a recipe once it scores around 9.5 out of 10.
Run past conversations through the agent and check each one lands in the right category.
Have a second model score every reply for accuracy and tone.
Run each test five times to check the scores are consistent.
Retest after every prompt or model change.
Rollout
04.
Hand over control in stages
A request type moves up a stage only after it has a track record at the one before.
Jp calls this charging a trust battery.
Cold
The agent writes an internal note for a person to review. The customer sees nothing.
Warm
The agent replies, and the conversation stays open for a person to check.
Hot
The agent replies and closes the conversation on its own.
Guardrails
05.
Build in limits
Two limits keep the agent in check.
Act only when sure
The agent handles a request only when it’s confident. Everything else goes to a person.
Resist manipulation
The agent is built to refuse when someone tries to talk it into doing something it shouldn’t.
Without these AI agents, support would always be playing catch-up.
JP Valery
Customer Success Engineer, Resend
Resend built two agents and brought them to Plain. One answers customers through Plain’s API and webhooks. The other is a local helper that uses Plain MCP to help with more complex threads.
5x
user growth in 2026
1 in 3
from 10% to 33% in 5 months
80%
automated resolution goal
Scope
01.
Start with one type of request
Pick a request that comes in often, gets the same answer every time, and needs data you can fetch easily. For Resend, that was domain verification.
Tip: Ask engineering for help early. Getting from prototype to production was Resend’s biggest delay.
Architecture
02.
Split the agent into three jobs
Resend calls each of these flows a recipe and runs more than 20 of them.
Understand
Classify the request. If the agent isn’t sure, a person takes it.
Gather
Fetch only the account data this request needs.
Reply
Write the reply from the message, the request type, the data, and your instructions.
Testing
03.
Test before customers see it
Resend trusts a recipe once it scores around 9.5 out of 10.
Run past conversations through the agent and check each one lands in the right category.
Have a second model score every reply for accuracy and tone.
Run each test five times to check the scores are consistent.
Retest after every prompt or model change.
Rollout
04.
Hand over control in stages
A request type moves up a stage only after it has a track record at the one before.
Jp calls this charging a trust battery.
Cold
The agent writes an internal note for a person to review. The customer sees nothing.
Warm
The agent replies, and the conversation stays open for a person to check.
Hot
The agent replies and closes the conversation on its own.
Guardrails
05.
Build in limits
Two limits keep the agent in check.
Act only when sure
The agent handles a request only when it’s confident. Everything else goes to a person.
Resist manipulation
The agent is built to refuse when someone tries to talk it into doing something it shouldn’t.
Without these AI agents, support would always be playing catch-up.
JP Valery
Customer Success Engineer, Resend
Resend built two agents and brought them to Plain. One answers customers through Plain’s API and webhooks. The other is a local helper that uses Plain MCP to help with more complex threads.
5x
user growth in 2026
1 in 3
from 10% to 33% in 5 months
80%
automated resolution goal
Written by
JP Valery
Customer Success Engineer, Resend
Customer Success Engineer at Resend
Written by
Susana de Sousa
Community, Plain
Head of Community at Plain
More guides
Playbooks and practical guides for B2B support teams.
Get your support running on Plain in minutes.
Get your support running on Plain in minutes.
Get your support running on Plain in minutes.
