The AI agent & Knowledge Hub

The feature is two things that work as a pair. The AI agent is the part that talks — it reads a customer’s message and replies in your voice. The Knowledge Hub is the part it knows from — a shared library of your facts. The agent doesn’t invent answers; it looks them up.

Both live in the agency area: AI Agents for the agents and Knowledge Hub for the library. One Knowledge Hub is shared across every agent, so a fact you fix once is fixed for all of them.

You can stop after three steps and have a real, useful agent. Here’s the whole minimal path before we go deeper:

A new agent starts as a Draft with a default model and a starter prompt. You won’t be able to use it in a workflow until you flip it to Published — that toggle is the line between “still editing” and “ready to run.”

Start with knowledge, not the agent — an agent with nothing to read can only fall back on generic answers. The Knowledge Hub has two tabs:

• FAQs — short question-and-answer pairs, grouped into folders (Pricing, Schedules, Policies…). This is the backbone; it’s where most answers should come from.
• Documents — files you upload (price lists, brochures, policy PDFs) for the longer, less structured material.

A FAQ is just a Question and an Answer. Write the answer exactly as you’d want the agent to say it — it’s used as the source of truth, near-verbatim.

When a message arrives, the agent doesn’t just free-associate. It calls a Search Knowledge Base action, pulls the most relevant FAQs and document chunks, then writes its reply from those. Each piece comes back ranked by how closely it matches the question.

Every AI reply can carry a “Searched N sources” pill. Expand it to see exactly which FAQs and documents the agent pulled, ranked by relevance, with the snapshot of text it actually read. If an answer looks wrong, you trace it to the source and fix the FAQ — and every future reply uses the corrected version.

Open an agent and you get an editor with a tab for each part of how it behaves. The one that matters most is the first: Instructions — the System Prompt that defines its personality, role, and how it should behave. This is the single biggest lever on reply quality.

The tabs, in plain terms:

• Instructions — the system prompt. Write it yourself or hit Generate with AI to draft one from a description, then refine it in plain English.
• Actions — the things the agent is allowed to do (covered next).
• Exit Conditions — the ways a conversation can end, and the reason it reports back.
• Guardrails — topics intercepted before the agent even sees them.
• Output Data — structured fields you want the agent to return (for branching later).
• Tests & Feedback — check it against saved cases, and review real replies.
• Settings — conversation limits and exit behaviour.

Actions are the capabilities an agent can use mid-conversation. They split into read-only data actions (on by default) and write actions you opt into. A write action quietly switches on the data actions it depends on — enable Create Lead and it locks on the “get pipelines / fields” reads it needs.

One action is special: Search Knowledge Base. It’s what grounds replies in your Hub, and you can scope it — All folders (the default) or Specific folders, so a sales agent and a support agent can read different shelves of the same library.

Every agent turn ends with an exit reason — returned as exitReason so the next workflow step can branch on it. Three are always there and can’t be removed:

On top of the defaults, add custom exit conditions for the outcomes that matter to your business — trial_booked, not_interested, needs_human. Each is a short key plus a description of when the agent should use it. Generate with AI can draft a sensible set straight from your system prompt.

Guardrails run before the agent reads a message. When a contact’s message matches a rule’s topic, it’s intercepted and a fixed action fires — the agent never sees it. That makes guardrails far more reliable than a “please don’t talk about X” line buried in the prompt. Each rule picks one of three actions:

Two limits in Settings stop a conversation running forever:

• Max Turns — each turn is one agent reply. Hit the cap (50 by default) and it exits with max_turns_exceeded.
• Session Timeout — optional; ends the session after a stretch of inactivity with timeout.

An agent does its real work inside a workflow. Drop an AI Agent node, pick your published agent and a channel, then branch on how it ended. That exitReason is the join between the agent and the rest of your automation.

The node has a couple of useful knobs. A Timeout caps how long it waits for a reply, and Burst message batches rapid-fire messages — if a customer fires off three texts in a row, the agent waits a beat and treats them as one, instead of replying three times.

Replies get rated with a thumbs up (Marked good) or thumbs down (Flagged). A flag asks for a category — Hallucination, Wrong tone, Missed context, or Other — and a note on what it should have said. Each flag lands in the agent’s feedback inbox.

From there it’s an admin loop, not an automatic one. Flagged items wait under Pending. When you hit Train agent, Exabloom bundles them into a proposed refined system prompt, shown as a before/after diff you review and apply. Items you’ve folded in move to Trained; ones that aren’t the AI’s fault you can Reject.

Four configurations, simplest first. Each is built from pieces above — adapt them, don’t copy exactly.

• Knowledge before agent. An agent is only as good as the Hub behind it. Seed your top FAQs first — pricing, schedules, location, policies.
• Publish to use. Workflows only see Published agents. A Draft is invisible in the AI Agent node’s dropdown.
• Fix facts in the Hub, behaviour in the prompt. Wrong figure → edit the FAQ. Wrong tone or structure → flag it and train.
• There’s no built-in “human handoff” exit. Use a custom exit condition you branch on, or Request Internal Input — they’re different tools.
• Scope the knowledge source per agent. Point each agent at the folders it should read so a sales bot doesn’t quote support-only policy.
• Training is an admin decision. A thumbs-down queues the item; nothing changes until someone reviews the refined-prompt diff and applies it.
• Turn off the exit message when chaining agents, so an intermediate agent’s closing line doesn’t get sent to the customer.