Guides5 min read

Tuning Your Agent's Behavior

Fine-tune how your agents communicate and make decisions — from the job description and system prompt to review mode progression.

The Behavior Tab

Every agent has a Behavior tab where you control how it thinks, communicates, and makes decisions. This is your primary tool for shaping agent output.

Job Description

Each agent has a plain-English Job Description that describes its role, purpose, typical activities, and expected outcomes. Symbi uses this as the top-level narrative for the agent and feeds it into the AI's context on every task.

A strong Job Description covers:

  • Role — what this agent is (e.g., "First-line customer support agent for billing questions")
  • Purpose — why the role exists (e.g., "resolve billing issues fast so customers don't have to wait")
  • Activities — the day-to-day work the agent does
  • Outcomes — what success looks like

Tip

Write the Job Description as if you were onboarding a new human teammate. The clearer your description, the better the agent performs — and the less tuning you'll need to do downstream.

System Prompt

The system prompt is the agent's core instruction set. From the Behavior tab, you can edit it to refine:

  • Tone and style — formal, friendly, concise, or detailed
  • Guardrails — topics to avoid, escalation rules, and boundaries
  • Specific instructions — edge cases, preferred response structure, or domain-specific terminology

The system prompt is sent with every interaction, so changes take effect immediately on the next task.

Classifier and AI-Assisted Categories

For agents that categorize incoming work (tickets, emails, messages), the Classifier controls how items are labeled. You can define custom categories with descriptions and judgment-based instructions, like:

  • Ticket priority — low, medium, high, urgent
  • Request type — question, complaint, feature request, bug report
  • Routing — which team or workflow should handle the item

AI-Assisted Generation

When you're setting up categories, Symbi can propose a starting set for you based on your company context and job description. Click Generate with AI to get a draft category list you can then refine, rename, or remove items from. It's the fastest way to get a solid baseline without starting from scratch.

Company Profile Influence

Your company's purpose, vision, values, and brand narrative (set in Settings) are automatically injected into every agent's context. All your agents share a consistent understanding of your brand — even without custom prompt edits.

Tip

The company profile is the single most impactful setting. A clear brand narrative improves every agent's output across the board.

Knowledge Base

The knowledge base gives your agent factual context to draw from. To get the most out of it:

  • Keep it focused — quality over quantity. A few well-maintained sources outperform a large, outdated collection
  • Prioritize FAQs — frequently asked questions and common scenarios should be front and center
  • Remove outdated content — stale docs cause bad answers
  • Update after process changes — whenever a policy or workflow changes, update the knowledge base

See Knowledge Base Management.

Review Mode Progression

One of the most powerful aspects of Symbi is gradually increasing an agent's autonomy. We recommend this three-stage progression:

  1. Start with Always Review — every action goes through human review. Builds your baseline understanding of how the agent handles different scenarios
  2. Move to Confidence-based (X%) — the AI auto-approves actions above your chosen confidence threshold; lower-confidence ones are still queued. Saves time while keeping you in the loop on edge cases
  3. Graduate to Autonomous — once you're confident, enable full auto-approve. You can still spot-check from the Operation Center any time

You can drop back to a stricter mode at any point — for example, during a promotion week when stakes are higher.

Per-Tool-Group Review for Agent-Mode Agents

Agent-mode agents support a finer level of control: each of the four action groups can be set to a different review mode independently.

| Group | What it covers | Common starting point | | --- | --- | --- | | Communicate | Customer-facing replies | Always Review | | Lifecycle | Ticket and record status changes | Confidence-based | | Organize | Categorization, tagging, prioritization | Autonomous | | Internal | Internal notes and comments | Autonomous |

A common starting point: keep Communicate on Always Review while setting Organize and Internal to Autonomous — so you approve every outgoing reply but don't need to review every label or note.

Configure these settings from the Behavior tab on the agent's detail page. The agent-wide review mode shown in that tab reflects the dominant setting across applicable groups.

Iterating Through Reviews

When you edit an action during review, you're not just fixing a single reply — you're teaching the agent. With Learn from Reviews enabled, edits become learned knowledge sources the agent references in future tasks. Over time you'll see:

  • Fewer edits needed per review cycle
  • More consistent tone and formatting
  • Better handling of edge cases the agent previously struggled with

This feedback loop is the fastest way to shape your agent's behavior without rewriting the system prompt.

What's Next?