SymbiSymbi
Getting Started6 min read

How Agents Work

Understand the architecture behind Symbi's AI agents — from templates and hiring to task execution, review, and learning.

The Agent Lifecycle

Every agent in Symbi follows a consistent lifecycle, from discovery to daily operation.

1. Templates

An Agent Template is a blueprint for a type of agent. It defines:

  • What the agent does (e.g., "responds to FreshDesk support tickets")
  • Which integrations it requires (e.g., FreshDesk + Slack)
  • Default behavior, tone, and review settings

Templates are curated by the Symbi team and appear in the Marketplace.

2. Hiring

You can hire an agent two ways:

Via the Marketplace: Browse templates, click Hire, and Symbi opens the new agent's page with a three-step checklist — give it apps, describe the job, check what it can do. You can leave and come back at any step; nothing runs until you press Publish & go live (v1).

Via your Personal Assistant: Describe what you need in chat. The assistant finds a matching template or asks a few questions to build one, then shows you a preview. During setup the builder guides you through picking the specific resources the agent will use — for example, which Zoho Desk queue to watch, which Gmail label to filter on, or which Slack channel to post to. If no existing resource matches, you can register one on the spot — paste a URL (SharePoint, Google Drive, or Zoho Desk), or type a Gmail label name (e.g., Support Tickets) directly. You finish on the agent's page with Publish & go live (v1) — publishing runs its own checks and refuses while a required resource slot is still empty or an integration is not connected.

3. Operating Agents

Once active, an agent belongs to your company and has:

  • A profile on the Agents page
  • Connected integrations (credentials stored encrypted)
  • Configuration for behavior, tone, review preferences, and knowledge
  • Performance metrics tracked over time in the Dashboard

How a Task is Processed

When something triggers an agent — a new ticket, a scheduled run, an incoming email — Symbi follows this flow:

  1. Event received — Symbi picks up the event from the connected service (webhook, polling, or schedule)
  2. Trigger filter — the agent checks whether this event matches what it's supposed to handle. For example, an agent may only respond to tickets with the "Billing" category
  3. Context building — the payload is enriched with your company profile, the agent's knowledge base, integration data (CRM records, previous messages), and relevant history. Attachments on the triggering item are included automatically, no extra tool needed: a Gmail email or Zoho Desk ticket brings its PDFs, images, Excel, Word, and text/CSV files (up to 10 per item), while a FreshDesk ticket brings images only. A hand-off from another agent carries the same files as the item it came from. Files Symbi can't read (zip archives, calendar invites, and similar) are skipped, and the agent is told which ones and why
  4. AI processing — the agent runs the task, choosing which tools to call and what actions to take
  5. Review gate — depending on the agent's review mode, actions are either auto-approved or sent to the Operation Center for human review
  6. Delivery — approved actions are sent back to the originating service (reply posted, status updated, note added, etc.)

Info

All actions an agent takes are logged and inspectable from the Activity tab in the Operation Center — you always have a full trace of what the AI did and why.

Triggers

An agent starts work when a trigger fires. Symbi supports three kinds:

  • Event trigger — an external system sends a webhook (e.g., new FreshDesk ticket, new email, Linear issue created)
  • Scheduled trigger — the agent runs on a time schedule (e.g., every morning at 8 AM, every 4 hours)
  • Manual trigger — you click Run Now on the agent's detail page to kick off a task on demand

Scheduled and manual triggers are ideal for agents that produce reports, do regular clean-up, or run recurring workflows. See Scheduled Triggers.

Human-in-the-Loop Review

The review system is central to Symbi. Every action an agent proposes can be:

  • Approved — delivered as-is
  • Edited — adjusted by a human before delivery
  • Declined — rejected, no action taken

Reviews happen in the Operation Center (in-app) or directly in Slack via interactive messages. Both stay in sync — approving in Slack updates the app, and vice versa. See Operation Center.

Knowledge

Each agent can be equipped with one or more Knowledge Bases — collections of documents, URLs, FAQs, and learned content that give the agent context about your business.

Knowledge is consulted in real time as the agent processes a task. You can add, update, or remove sources whenever you need. See Knowledge Base Management.

Auto-Extending Your FAQ

Each agent has an Auto-extend FAQ from support tickets toggle in the What it does section of its page. When enabled, a daily distillation run analyzes the agent's approved and edited ticket reviews since the previous run (with a 24-hour overlap, so nothing near the boundary is missed), extracts FAQ patterns, and adds them to the agent's knowledge base automatically. Your manually uploaded content stays authoritative — the auto-extension supplements it without overwriting anything.

This creates a feedback loop: the more you review and refine, the richer the agent's FAQ becomes. Over time, most teams see fewer edits needed per review cycle and better handling of recurring scenarios.

Performance Tracking

Symbi tracks key metrics for every agent:

  • Interactions — total tasks handled
  • Success rate — percentage of actions approved without edits
  • Hours saved — estimated time saved vs. manual handling
  • Cost saved — financial value of those hours

These metrics surface on the Dashboard and on each agent's detail page. See Performance and Analytics.

What's Next?

Previous

Connecting Slack

Next

Operation Center