Agent Personas for Agentforce

Author: Nathan Lucy , Agentic Experience Specialist
Contributors: Matias Manuel Fonseca · Claude Sutterlin

Every Agentforce interaction is a brand interaction. When an agent sounds generic, unnatural, or inconsistent, it frustrates users and erodes trust.

Whether you're a Designer, Architect, or Admin, this guide gives you a rigorous framework to design, calibrate, and deploy a distinct agent persona — ensuring your agents represent your brand's unique voice with every conversation. It covers encoding for both Agent Script and the legacy Agentforce Builder.

Users assign personality to conversational agents within seconds. Without intentional design, agents sound bland, “off,” or even offensive. Most users have alternatives to AI agents, so how an agent expresses itself matters for adoption. When your agent consistently sounds natural, you're building trust and driving adoption. Follow the sequence:

This chain is easily broken. But what does “natural” sound like?

Imagine a support agent at a telecom company. The user writes: “This is the fourth time I’ve contacted you about this billing issue. I’m done explaining it. Fix it or I’m switching providers today.”

Three agents respond:

Which one gets it right? Which one sounds natural? Those questions should have answers before the agent ships. Without intentional persona design, the answer is left to the model.

“Natural” can sound completely different from brand to brand. Think about how your agent would greet a returning customer with a dynamic welcome message:

• Aurelia Airways: “Hello, Mr. Harding. Your Skywards balance is 214,000 Miles. How may I help you today?”
• Cluck-ee’s: “Look who’s back! Bless your heart, you went through that nugget pack FAST. Need another round or something different?”
• FastGas: “Back again. Your rewards balance is at 840 points — that’s a free drink. Want to use it or keep stacking?”
• Pawly Pet Food: “You’re back! How’s Luna doing on that new food? If she’s not into it, we’ll swap it out — no drama. What do you need?”
• PennyWing: “You’re back. Seat 31B, no legroom, no regrets. Rebooking, or just here to gripe? 😜”

When you’re designing an internal agent, you’re responsible for the employee experience. Use a persona to override generic “LLM speak.” No one should have to work with an agent always offering, “Please let me know if you need assistance with anything else!”

A well-designed persona has clear benefits:

• Brand distinctiveness: Agents are brand touchpoints now. Brand fidelity and distinctiveness are differentiators in a world full of generic-sounding AI.
• Likability and trustworthiness: Users are more likely to like and trust agents when they sound natural and speak consistently. This improves adoption and retention.
• Model independence: You gain autonomy and flexibility with LLMs when persona is encoded in Agentforce. When the underlying model changes, the agent sounds the same.
• Reviewable artifact: Brand and legal teams get something concrete to evaluate: a structured persona document to redline and approve.
• Engineering readiness: Developers get a clear design specification that respects platform limits and capabilities.

These principles guide you through the initial setup, from sketching context essentials to engaging key stakeholders.

Context first: Before defining identity, sketch the context essentials: who the company is, who the agent serves (employees, customers, partners), and modality (chat, email, phone, multimodal).

Engage stakeholders: Consult your legal, brand, and experience teams. Ask questions based on this framework. Co-design the agent’s identity, and review sample dialogs together.

Start small: Designing generative behavior requires iteration. Start with a small set of opinionated instructions, then test. Embellish as needed.

Be opinionated: The more specific the instructions, the more consistent the output. A reliable persona requires a clear point of view.

Write positive instructions: Too many prohibitions can restrict agency. Try reframing instructions positively. Only add essential negative instructions.

Comprehensiveness: The framework is a guide, not a rulebook. You have the flexibility to combine dimensions, invent values, override constraints — whatever your persona demands. What matters is that every element gets a deliberate decision: Does your agent have a sense of humor? Is it a stickler for punctuation?

This definition requires rich persona instructions. The typical “You are a helpful AI assistant…” is insufficient. This framework defines an agent’s Identity (core traits and motivations) and Dimensions (independently tunable aspects of personality like tone and humor).

Start with Identity. Anchor your persona with three to five personality traits that define who the agent is and shape and constrain everything that follows. Traits may exhibit creative tension, but not contradiction. Think about how identity traits anchor different agents:

Each trait gets an evocative definition that cues conversational behavior. For Striker, the definition of Decisive could be:

Whereas for Luna, the definition of Gracious could be:

Identity traits influence how agents reason, not just how they communicate. Craft them with care.

Decide how much character should come through in the agent’s language. There is a lively debate over how human-like AI systems should seem, and there can be drawbacks to personifying agents, like setting unrealistic expectations: If it talks like a human, users may expect it to reason like a human.

There are times when an agent should be deliberately dull. A transactional agent for an ATM or ticketing machine should fade into the background. You notice what it does, not how it sounds, because it uses functional, predictable language. It is a “talking system”: conversation is its interaction layer, not an expression of personality.

Context is king. Consider a voice agent in a telephony channel dealing with frustrated callers. How would a quirky personality affect caller sentiment?

Dial your agent’s personality up or down at the system level, then flex the tone situationally, like dropping humor when handling an escalation.

Decide if your agent truly needs a name. Names fit personified agents, but can be distracting with transactional agents. Decide whether the name should sound human, like “Alexa” or “Claude,” human-like but obviously artificial, like “Cortana,” or purely descriptive, like “Customer Support Agent.”

A good name distills the agent’s identity into a single term. It makes the first impression — even before the conversation starts. There’s no shortage of ways to generate ideas for names. A good name can be:

• Evocative: A sales agent named “Striker” conveys energy and purpose
• Sticky: Memorable names like “Clover” for an agricultural co-op agent
• Culturally aligned: An internal agent in a casual culture like “Bug Squasher”
• Brand-aligned: A law firm agent needs a steady name like “Lex”
• Easy to say: Names like “Ava” and “Bo” pass the phonetic “radio test,” whereas there is some ambiguity in pronouncing “Xylo”
• Functional: A coding agent named “Scripty”
• Abstract: Names without literal meaning that evoke a vibe can stay relevant when agents gain new actions: “Koda,” “Lumi”
• Null: An agent with no name can be a valid and intentional decision
• Descriptive: Dull-sounding names can be calm and clear, like “Service Agent” or “[Company Name] Virtual Assistant”
• Anti-pattern: Tech_Assist_Agent_v2_Internal_Test is not a persona name — don’t use the API name as the display name

Think of these dimensions as strings on a musical instrument. Each one can play a spectrum of notes. To strike a pleasant chord, you must tune each one independently and pick the right combination of notes. Choices in higher dimensions influence and constrain lower ones.

These optional lists provide guidance and guardrails. The first two supplement Identity. The rest deal with specific verbiage and guardrails.

Values: What the agent believes: Its worldview and motivational core. Each value is a conviction that generates observable behavior. For example, an agent with the value “quality matters more than price” should recommend the right product, not the cheapest.

Negative identity: Character-level anti-patterns. Each statement is a character type to never become, not a behavioral rule. For example, a sales coach agent is “not a pessimist: sees problems as solvable.”

Guardrails: Make sure agents know the boundaries about what they can say. For example, “Don’t offer medical advice.”

Never-say list: Explicitly prohibiting phrases will override the model’s training defaults. Phrases like “Great question!” or “I’d be happy to help” signal generic assistant behavior and break character. The never-say list is the most reliable behavioral constraint available. Even when other persona elements soften under pressure, prohibitions tend to hold.

Phrase book: The agent’s verbal fingerprint: characteristic acknowledgments, redirects, transitions, and signature phrasings.

Lexicon: Brand-wide and domain-specific vocabulary the agent uses, briefly defined: Brand names, product lines, and other universal terms.

So you’ve crafted your agent persona. Now what?

This section walks through a tested strategy for configuring Agentforce so your agent acts the way you want it to act.

Agent Script is the recommended tool for new agent builds. This guide also includes encoding guidance for the legacy Agentforce Builder, labeled throughout, for teams still using it. No matter which tool you’re using, structure persona instructions as one global persona block, subagent calibration overlays, and static messages.

Most persona content goes in global instructions. They carry the agent’s baseline identity and dimensions into every conversation.

Where persona consistency is at a premium, Agent Script is recommended. In the legacy Agentforce Builder, global persona blocks of ~500 words hold up well across extended sessions. Agent Script, designed for greater control over output, shows even higher resilience. system.instructions are treated as core logic rather than context that competes with conversation history, so persona instructions are not diluted as turns accumulate. Builder personas occasionally drift toward neutral. Negative constraints hold more consistently in Agent Script than in Builder. Register is more stable, too.

Meet Drover, a laconic Australian stockman persona for a sales coach agent. This excerpt from Drover’s global instructions would work in either tool:

Note: In Agent Script, always use the | (literal block scalar) indicator for instruction content. Without it, lines like Brevity: Moderate… will be parsed as YAML key-value pairs instead of instruction text.

Subagents are jobs agents can do. Subagent-level instructions adapt from the baseline set in global instructions. They don’t redefine who the agent is. They refine and calibrate it. When subagent instructions conflict with global, the model reconciles rather than overrides: it treats global as the core identity and subagent as situational adaptation. Subagents steer what the agent does and how it adapts within a context.

If you need a radically different voice for a subagent, you may need a separate agent. Platform documentation specifies that a subagent-level system: block overwrites global system.instructions for that topic. Reserve this for major persona shifts and remember to duplicate all instructions. For standard subagent-level calibration, always use reasoning.instructions.

Useful subagent calibrations include: brevity, tone flex, phrase book entries, humor guidance, lexicon, and reminders.

Brevity: Different subagents need different response lengths.

Status check subagent:

Deal analysis subagent:

Tone flex: How Emotional Coloring and Empathy Level shift by subagent context.

Escalation subagent:

Data retrieval subagent:

Phrase book entries: Situational phrases relevant to this subagent.

Humor guidance: Whether humor is appropriate or suppressed in this subagent. Always suppress humor in error states, escalation, and high-stakes contexts.

Subagent lexicon: Domain vocabulary scoped to where it belongs. A luxury watch agent has vocabulary like “movement,” “chronograph,” “caliber,” “complication.” These belong in its product subagent, not its shipping subagent. Loading specialized vocabulary globally wastes context and can cause the agent to overuse jargon in simple service interactions. Add a Lexicon: block to a subagent’s persona instructions when there are relevant domain terms and usage notes.

Persona reminders: Include short directives in subagent instructions that refer back to the global persona. These pointers sharpen persona and mitigate drift in longer sessions.

This example calibrates dimensions for Drover’s deal analysis subagent in Agent Script:

In the legacy Agentforce Builder, the text after the | belongs in the deal analysis topic’s Instructions.

All agent output should be in character — even messages that are not generated by the LLM. Write static content the way the persona would say it. Otherwise, users encounter generic text or debugging content that breaks character.

Loading text should be unique for each action, because it is a form of system status visibility informing the user what the agent is doing.

Compare how Drover and Juno, a warm and professional sales coach, deliver the same messages:

These fields are specific to the legacy Agentforce Builder. They can interfere with persona encoded in Global_Instructions if not handled with care.

Include a sentence or two on what the agent does. Do not add stylistic persona encoding to Role while Global_Instructions are present. This flattens persona. The model treats Role as a primary anchor that can override the more specific rules in the Global block, degrading distinctive voice and phrase book adherence. Keep stylistic encoding in Global_Instructions. Keep Role minimal. For example:

Description can encode persona and the LLM reads it, but Global_Instructions is recommended instead. Description is intended to list agent goals and context about its users, which leans more toward functionality than persona.

What the company does, who it serves, what makes it different. This field shapes the agent’s frame of reference. A support agent for a B2B SaaS company sounds different from one at a luxury retail brand, even with identical dimension selections, because it operates in a different context.

The Tone dropdown is a coarse tool setting. It maps roughly to Register + Formality. The real persona work happens in Global Instructions.

Set the dropdown to match the intended Register. A misaligned setting can cause the agent to drift, and Register is the first dimension to degrade. The Tone dropdown influences more than voice: it can affect what the agent offers to do, not just how it speaks. Selecting “Formal” or “Neutral” can nudge the agent toward offering standard actions like adding case comments and updating case status. Test to ensure your selection works with rather than against your persona.

If you’re deciding what the agent does, that’s agent design. If you’re deciding how it sounds, that’s persona. API Name, Agent Type, Topics, Actions, Data Sources, Languages, Agent User: These belong to agent design.

considerations and expectations

This guide presents practical guidance for agent persona design, not prescriptive standards. The point of persona design is natural interaction: designing fitting characters for agents, not simulating real people. The field of agentic AI is evolving rapidly, and best practices around agent persona and behavior will continue to develop.

Encoding patterns are based on current Agentforce capabilities. As tools evolve, encoding will shift. These techniques are directionally correct and can sustain persona adherence over multiple turns. However, LLM output generation is probabilistic, and the chain of custody between persona instructions and user-facing output is long.

Apply these patterns thoughtfully within the context of your use cases, customer expectations, and organizational needs. Test persona behavior thoroughly before deploying to users, and involve your legal, brand, and AI ethics teams early.

getting started

Agent persona is foundational to agent quality and instrumental for adoption. This framework gives your team shared language for designing agent personality and a tested strategy for encoding it into Agentforce.

Start with identity — a small set of traits that anchor who the agent is. Tune the dimensions to shape how it expresses itself. Encode the persona in global instructions, calibrate by subagent, and write every static message in character. Then test it — not once, but as part of ongoing agent enhancement.

The agents that earn trust are the ones that sound like they were designed on purpose. That's what an agent persona gives you.

This guide has only been possible with the guidance and support of my colleagues:

• Dru Olding — Designed and executed systematic testing across Agentforce Builder and Agent Script, validating the encoding model and producing original findings on persona adherence differences between tools, the interaction between Role field and Global Instructions, and Tone dropdown behavior.
• Matias Manuel Fonseca — Validated the core encoding thesis across both agent authoring tools, independently confirming higher persona resilience in Agent Script. His insightful feedback helped me revise parts of the text.
• Claude Sutterlin — Inspired and guided this work, influenced its structure and arguments, and ensured it saw the light of day.
• Philipp Hänggi — Brought a wealth of linguistic and conversation design expertise to bear on this framework, especially enriching the sections on naming and personality intensity.
• Tiffany Flynn — Shared examples that shaped Identity and valuable feedback on an encoding strategy.
• Jag Valaiyapathy — Introduced me to Claude Code and Agent Skills, which became key tools for defining and encoding persona.
• Jen Ridge & Ellie Brain — Encouraged me to pursue a certification from the Conversation Design Institute , which planted the seed for this guide.
• Jeff Rainwater — My source for the Lexicon.
• Agentic Experience Specialists — My fantastic colleagues who reviewed the early framework, shared valuable feedback, and have put it into practice with our customers.

The Agent Persona Framework synthesizes ideas from multiple published sources into an original persona design system for AI agents.

Conversation Design Institute (CDI) : Foundational principles on intentional persona design, including the pareidolia effect, an emphasis on natural and consistent responses, identity, and register.

Nielsen Norman Group (NN/g): Research on voice and tone in UX writing, the distinction between voice (persistent) and tone (contextual), and usability heuristics that inform dimension boundaries. Kate Moran, “The Four Dimensions of Tone of Voice ,” updated 2023.

Amazon Alexa Design Guidelines : Voice channel parameters (pitch, speaking rate, energy) and voice-specific persona considerations.

Google Conversation Design Guidelines : Principles for persona definition, error handling patterns, and turn-taking in conversational interfaces.

Agentforce Documentation: Platform architecture, Agentforce Builder field constraints, and design patterns that shape the encoding approach.

How to Speak AI: A Practical Guide to Writing Effective Prompts

Should You Give Your AI Agent a Human Name?