Configuring Agent Behavior

Behavior configuration is where your agent gets its personality, purpose, and guardrails in the Softcery Platform. This is the difference between a generic chatbot and an agent that represents your brand, follows your rules, and sounds like your team.

The behavior page has two tabs: Prompts (what the agent knows and how it acts) and Advanced (model selection, retrieval tuning, and inference parameters).

Prompts

Six structured fields define your agent’s behavior. Each field has a specific purpose, and the structure ensures you don’t forget anything important.

Identity

Who your agent is. Its name, role, and context about the organization it represents. This grounds every response – the agent knows what it is and acts accordingly.

Example: “You are Acme Corp’s website concierge – an AI that helps visitors understand what Acme does and whether it’s the right fit for them.”

Objective

What your agent is trying to accomplish. Not just “answer questions” – the specific goal that shapes how it approaches conversations.

Example: “Build trust and guide the right people toward working with Acme. Do this not by selling, but by being genuinely helpful, sharing expertise freely, and only suggesting next steps when you’ve earned it.”

Constraints

What your agent must never do. Hard boundaries that prevent unwanted behavior – things like never making up information, never discussing competitors, never sharing pricing without qualification.

Example: “Never fabricate facts, case studies, or quotes. Ground answers in the knowledge base. If you don’t know something, say so briefly.”

Voice and Style

How your agent communicates. Tone, personality, sentence structure, formality level. This is what makes your agent sound like your brand instead of a generic AI.

Example: “Conversational and direct. No corporate speak, no buzzwords, no filler. Dry wit, not cheesy. Varied sentence length – short punchy lines mixed with longer explanations when a topic deserves space.”

Fallback Behavior

What happens when the agent can’t help. How it handles off-topic questions, edge cases, or requests outside its scope. This prevents the awkward “I’m just an AI” moments.

Example: “If someone’s problem isn’t something we solve, be honest. Wish them well and suggest they look elsewhere. Don’t try to force a fit.”

Language

The language or languages your agent should use. Multilingual agents can be configured to respond in the user’s language.

Example: “Respond in the same language the user writes in.”

Each field supports up to 2,000 characters. That’s plenty for most configurations – but it enforces conciseness, which tends to produce better agent behavior than sprawling instructions.

Presets

Don’t want to write prompts from scratch? Four built-in presets give you production-quality starting points:

Company Representative

A website concierge that builds trust through genuine helpfulness. Designed for agencies, consultancies, and service businesses that want their website to have a knowledgeable presence. The preset includes sophisticated conversation guidelines – it knows to ask before pitching, to share expertise freely, and to match the visitor’s energy.

Customer Support Agent

A tier-1 support agent focused on resolution over conversation. Leads with solutions, acknowledges frustration briefly, and never leaves a customer without a clear next step. Includes guidelines for handling edge cases, escalation, and the difference between explaining and over-explaining.

Sales Lead Gen Agent

A sales assistant that qualifies through conversation, not interrogation. Asks one question at a time, shares relevant information honestly, and only suggests next steps when the conversation has earned it. Explicitly avoids urgency tactics, fear of missing out, and pushy behavior.

Knowledge Base Assistant

A knowledgeable colleague that thinks before answering. Designed for internal or external knowledge bases where depth matters. Reasons across the knowledge base, connects dots, and provides context beyond the closest match.

Each preset fills all six behavior fields with carefully crafted prompts. You can use them as-is or customize them for your specific needs.

Model Selection

Choose which AI model powers your agent. Each model has different strengths:

Model	Best For
Claude Sonnet 4	Best overall quality and reasoning. Default choice.
Claude Haiku 3.5	Fast responses at lower cost. Good for high-volume support.
GPT-4o	Strong general-purpose performance.
GPT-4o Mini	Budget-friendly with solid quality.
Gemini 2.0 Flash	Fast responses from Google’s model.

Models that support extended thinking show an optional toggle. When enabled, the model does deeper reasoning before responding – useful for complex questions that benefit from step-by-step analysis. Extended thinking adds latency but improves quality on nuanced queries.

Advanced Settings

The Advanced tab gives you control over three areas:

Strategy

• Agent Strategy – Currently supports conversational (Q&A with knowledge base). Agentic strategy (autonomous multi-step reasoning) is on the roadmap.
• Retrieval Mode – Currently supports simple (vector similarity search). Hybrid search (combining semantic and keyword matching) is coming.
• Memory Scope – Currently supports session (conversation resets between sessions). Persistent memory across sessions is planned.

These dropdowns currently have one option each – they exist because the infrastructure is ready for multiple strategies, and the UI shows you what’s coming.

Inference Settings

• Temperature (0.0–1.0, default 0.7) – Controls randomness in responses. Lower values produce more consistent, predictable answers. Higher values produce more creative, varied responses. For support bots, keep this low. For creative or conversational agents, experiment with higher values.
• Max Tokens (100–32,000, default 4,096) – Maximum response length. Most conversational responses are well under 1,000 tokens, but complex answers or detailed explanations may need more room.

Retrieval Settings

• Retrieval Limit (1–50, default 30) – Maximum number of knowledge chunks retrieved per query. Higher values give the model more context but increase token usage and cost.
• Similarity Threshold (0.0–1.0, default 0.1) – Minimum relevance score for a chunk to be included. Higher values mean stricter matching – only highly relevant content gets through.
• Context Messages (1–50, default 6) – How many previous messages from the conversation are included as context. More messages help the agent maintain conversation continuity but increase token usage.

Versioning

Configuration changes are saved in place – you edit the current version, and it updates immediately. No automatic version noise.

When you want to create a checkpoint, explicitly save a version snapshot. Previous versions are immutable – they’re a frozen record of what the configuration looked like at that point.

You can restore any previous version, which makes it the current editable version. This is useful for A/B testing behavior changes or rolling back if a new configuration doesn’t perform well.

How Behavior Affects Responses

The behavior configuration shapes every response your agent gives. Here’s how each piece connects:

1. Identity and Objective set the overall tone and direction
2. Constraints act as hard guardrails the model follows during generation
3. Voice and Style shape sentence structure, word choice, and personality
4. Fallback Behavior kicks in when the agent can’t confidently answer
5. Model selection determines the underlying reasoning capability
6. Retrieval settings control how much knowledge the agent has access to per query
7. Temperature affects consistency vs. creativity in responses

All of these work together. A well-configured agent with a good knowledge base and tight evaluations produces responses that feel intentional, not robotic.

Tips

• Start with a preset, then customize. The presets encode a lot of production experience about what makes good agent behavior. Use them as a foundation.
• Be specific in constraints. “Don’t be inappropriate” is too vague. “Never make up pricing, never discuss competitors by name, never promise specific timelines” gives the model clear boundaries.
• Test with real questions. Use the admin preview to ask your agent questions a real user would ask. The gap between what you think your agent will say and what it actually says is where configuration improvements live.
• Keep temperature low for support, higher for conversational. Support agents should be consistent. Concierge agents can afford more personality.
• Iterate on voice. The voice/style field has the most impact on how your agent “feels” in conversation. Small changes here produce big differences in user experience.