Teaching an agent with corrections
The same Inbound Leads Agent that qualifies real email and WhatsApp messages can also be corrected and drilled on messages that never happened, and both paths teach it the same way: by writing a rule the agent checks against every future message, on either channel.
One agent, assigned to more than one connector
Why the same agent handles both channels in the first place: on the Connectors
page, each connected inbox or channel has its own Agents list. The
Inbound Leads Agent is switched on separately for IMAP Email and for WhatsApp,
the same agent record in both places rather than two copies of it, so it's one
set of qualification rules and one growing set of playbooks behind whichever
channel a message happens to arrive on.
Correct a real decision, or invent one that never arrived
On the Messages tab, every parsed IMAP and WhatsApp message shows its
confidence score and reasoning next to a Correct button. On the
Training tab, there's no need to wait for a real message at all: a
Mock inbound message form lets someone pick a connector, fill in
a from-address, sender name, subject, and body, and hit Analyze
to see exactly what the agent would have decided, without anything actually
having arrived.
A training chat turns a correction into a rule
Either path lands in the same training chat. A person describes the right outcome in plain language, for example that "need a ride to the airport" should count as a lead for a rideshare business even though it's short and informal, and the agent proposes a concrete rule back for confirmation before anything is saved.
The rule is remembered and reused, on either channel
Once confirmed, the correction is written as a versioned playbook entry tied to the agent, not to the message's connector. The next inbound message, whether it arrives over IMAP or WhatsApp, is checked against that playbook before the agent decides, which is why a rule taught from a WhatsApp correction ("Need car to airport") was later seen applying to a different WhatsApp message about a ride to the CBD, and why the same mechanism works for a mocked IMAP message just as well as a real WhatsApp one. What persists is the instruction text injected into future prompts, not model weights, and only confirmed corrections create a rule: an agent that runs unattended without anyone correcting it learns nothing new.