Under the hood · How Sidekick thinks

Never wonder why the AI said that.

Every Sidekick conversation is run by eight AI specialists reasoning over one shared conversation: reading intent, weighing sources, planning hand-offs. No script, no decision tree: judgment at every turn. Here’s one conversation, thought through live.

One assistant, a team of specialists.

Instead of one model guessing at everything, each decision in a conversation goes to a specialist built for that exact job, and all of them share the same conversation.

Specialized

Each agent is built and evaluated for one job: understanding intent, choosing a response, writing an answer, gathering ticket details, and so on.

Context-aware

One shared conversation flows through every agent, so decisions stay consistent from the first message to the last.

Matched, not one-size-fits-all

The system calls on different AI models for different jobs, choosing the one best suited to that specific decision.

The eight agents, at a glance

Each one owns a single decision in the life of a conversation.

1. Understand
   Routing Agent
   Understands & routes
   Reads every new message: what's really being asked, and how they feel, then routes it to the right specialist.

2. Route
   Knowledge Agent
   Selects the response
   Searches your knowledge, then picks exactly one way to respond: not just an answer, the right kind of answer.

3. Respond
   Synthesis Agent
   Composes the answer
   Writes a fresh answer grounded in your own knowledge base, and cites where it came from.

4. Hand off
   Follow Up Agent
   Gathers what's missing
   Requests the details a support ticket will need, and knows when it has enough to hand off.

5. Hand off
   Ticket Agent
   Ticket preparation
   Turns the whole conversation into a clean, correctly-routed ticket the desk can act on.

6. Respond
   Catalog Agent
   Service catalog
   Guides customers through your organization's official service requests, field by field.

7. Hand off
   Escalation Agent
   Reads the reply
   Reads how a customer responds to an offered next step: a yes, a no, or a hesitation.

8. Respond
   Web Search
   Live web fallback
   Looks beyond your knowledge base for trusted sources, only when your organization allows it.

A roundtable, not an assembly line.

All eight agents sit around the same live conversation: every message, every detail gathered, every decision made so far. There’s no fixed pipeline: the specialist the moment calls for takes the floor.

One shared conversation

Every message, every detail gathered, every decision so far: read by all eight, all the time.

Nothing is repeated

What one agent learns (a detail, a failed fix, a mood), every agent knows. The employee never says anything twice.

The floor changes hands

A question can become a request, then a ticket, mid-conversation. Whoever’s specialty fits the moment takes over, with the whole story.

Judgment, not scripts

No decision tree. Each turn, the team re-reads the conversation and decides what this moment actually calls for.

Matched to the job

Different decisions call for different strengths, so the system doesn’t force one model to do every job.

Fast, precise classification

Understanding intent and choosing a response path are quick, structured decisions, handled by models built for speed and consistency.

Nuanced writing

Composing a clear, well-grounded answer calls for a model strong at natural, careful writing.

Careful reasoning

Reading a customer’s reply, gathering ticket details, and structured hand-offs draw on models suited to step-by-step reasoning.

The system genuinely draws on more than one leading AI provider, choosing the model best suited to each individual decision rather than routing everything through a single one-size-fits-all model. That choice is an implementation detail your team can tune per agent as models improve, the eight decision points above stay the same.

The same principles, every agent, every time

Being built from specialists doesn’t mean being unpredictable, every agent follows the same ground rules.

Answer from what’s actually there

The Synthesis Agent and Web Search only speak from documents or sources that were actually retrieved, never open-ended invention.

Stay inside defined boundaries

The Catalog Agent and Ticket Agent only offer services and categories your organization has actually defined.

Ask rather than assume

When information is missing, the assistant asks a focused follow-up instead of guessing at an answer or a ticket’s details.

Know when to hand off

Clear signals (an explicit request, frustration, or a stalled conversation), route a customer to a person rather than looping.

One shared conversation

Every agent works from the same context, so a customer is never asked to repeat something they already said.

Honest about not knowing

When nothing available answers the question, the assistant says so, rather than presenting a guess as a confident answer.

Your assistant isn’t one model guessing.

Eight specialists, one conversation

Each agent owns a single decision, and all eight share the same context, all the time.

Every reply is a judgment call

Every possible way to respond gets weighed before a single word is written.

Grounded by design

Answers come from what was actually found: cited, never invented.

Built to hand off

Tickets, service requests, and live agents are first-class outcomes, not fallbacks.

Matched to the job

Different decisions draw on different AI models, chosen for what each does best.

“Not one model guessing: a team of specialists, each doing one job well, sharing one conversation.”