Agent Behavior Is the New Conversation Design

Hard-won patterns for anyone shipping multilingual agents at enterprise scale

Merci ou Merci bien 😉 ?

A wink emoji can make or break customer trust. When an agent signs off with “Merci bien 😉,” half of French users may smile, lightly, while the rest… may not. Spanish, Portuguese, and Japanese offer the same minefield: “Muchas gracias” signals warmth, “Muito obrigado” adds formality and gender, “Arigatō ございます” jumps a register. One byte of politeness often separates delightful from dismissive. Behavior is the product.

Why Behavior Beats Instructions

Early in enterprise AI we treated prompt snippets like final UI copy: tweak a noun, ship to prod. Then, imagine this piling on: brand-voice revisions, legal constraints, and forty languages. The result? Prompt-instructions collide, tool calls tangle, and trust trembles. Conversation design really begins at runtime, when behavior must adapt at every turn.

A way forward? A playbook grounded in field data & a few tails from battle.

Six Principles from the Trenches

1. Prioritize the Biggest Unblockers

Great roadmaps rescue users, not egos. Let me give you an example that is close and dear to me. At Agentforce, launching these four waves: English tooling first, FIGS-J-P responses second, AI-assisted authoring third, and full multilingual authoring plus social-skills tuning on its way to last. What’s the goal? That each phase unblocks a fresh customer cohort before edge cases spread. The lesson: release the capability that frees the most builders tomorrow, not the flashiest demo.

2. Data-Prep Beats Prompt Brute-Force

If Japanese RAG accuracy stalls at 70 %, do yourself a favor and resist prompt gymnastics. Instead, re-chunk docs, swap in a Jap-tuned embedder, and rerun thousands of evals. Then see accuracy leap to 90 %. Clean data plus the right retriever and embedding-model will frequently outperform ever-longer prompts.

3. Safety & Empathy Share an Architecture

Enterprises demand speed and trust in equal measure. Here’s an approach to support this. We’re shipping a single config-file that bundles tone rules, guardrail-topics, small-talk, slot-filling, planning-and-reasoning prompts, and tool calls, strategically, into LLM requests. For instance: a classifier routes queries to intent-specific “topic bundles”, a watchdog of sorts flags jailbreaks attemps, and, escalations hop to humans or other specific AI-agents. Result: security reviews halved and in customer-support contexts, first-contact resolution is rising beyond 20%. Empathy is safer, not slower, when it rides the same architecture-rail as governance.

4. Audit Trails Are UX

You are probably hearing this from customers, that they are asking or requiring: observability, explainability, and not a black box type of system. Me too. How to tackle it? Build “replay viewers” or reasoning-tracers that show every hidden chain-of-thought, tool call, and guardrail-topic decision. Auditors stop guessing, authors spot regressions fast, and users will gain a shareable “receipt” of what happened. Expose those knobs. Visibility nutures iteration; iteration nutures accuracy.

5. Build UI That Builds Itself

What if Admins could now ask, “Show me a churn dashboard,” and watch code render, and regenerate when they tweak a filter? At that point, the underlying agent would decompose the request, backtrack if needed, and surface reasoning for clarity. This is already here. That thing now called vibe-coding and its flip side. From where I’m standing, we call it Next Generation Authoring. As long as the system is “typed”: anything can be AI-assisted. Invisible software is closer than it looks and customers expect it.

6. Determinism When Stakes Are High: Agent Graphs

Talk to your top-100 customers. What are they saying? Please share! In the enterprise realm (but not only!), predictablily is at the rendez-vous for billing, returns, upsells, discounts, insurance, healthcare flows and more. By embedding deterministic agent graphs, which are explicit nodes, edges, and fallbacks inside the GenAI-reasoner, we blend LLM flexibility with predictable safety nets. For instance, a Fortune 500 insurer, may finally be able to cut policy-flow errors by 40% after graph-gating the overall experience. And these hybrid experiences keep auditors happy and users pleasantly surprised.

With core principles in place, the next hurdle isn’t technical, it’s relational.

Human — AI Bonds: Perceived Consciousness as a Design Problem

Friends, family, and fools (kidding!) keep telling me, “ChatGPT feels like someone.” I even wrote a plain-language primer for film students in april this year — Demystifying LLMs for Storytellers — and I feel the feeling persists. Joanne Jang’s excellent essay echoes this in “Some Thoughts on Human-AI Relationships.”

We can’t resolve “existential consciousness” today, but we can engineer for perceived consciousness, just as we already manage perceived latency. When an agent says, “🔍 Fetching invoices…,” eight seconds feel like three. Conversation Designers, arguably the next generation of elite prompt engineers, apply the same tactic to empathy: reflect intent, paraphrase clearly, reveal reasoning cues. Perception, not physics, defines the user’s reality.

That brings us to the balancing act: be polite without inventing a soul.

Polite, Not Personified

Said differently, the goal is to provide a system that’s respectful and restrain from inventing or infusing to it a dramatic personality. Does that make sense? The way I’m approaching it as I’ve been contributing to building the Agent-Plaform with its agent-behavior configuration capabilities at Agentforce from Salesforce is by shipping it with out-of-the-box “Small-Talk” topics so builders can:

• Use it immediately for greetings and gratitude.
• Retune tone to fit brand voice — eg. cheery retail vs. formal banking.
• Disable it altogether when regulatory context demands zero joking.

The rule: friendly language without fabricated feelings. No invented fears, crushes, childhood memories or survival instincts: just clear, courteous context. Approachable, yes. Anthropomorphic, maybe not.

Where We’re Headed

Once empathy and guardrails share a backbone, the obvious question is what’s next? Here are the frontiers my team, and the market, are racing toward.

• Chain-of-Thought RL: will let agents tackle for example medical or legal workflows, recover from dead ends, and expose reasoning for audit.
• Generative UI: is and will be about turning specs into live dashboards, stories, and even generating tools, mini-agents, and apps on demand at runtime {vs. at config-time, rn}.
• Behavior Specs: lintable files that include guardrail topics and deterministic graphs will ship alongside every API release.

We’re building toward an era where “Behavior Specs” sit beside Open API specs in every repo.

Call to Action

If you’re wrestling with multilingual tone, emotional perception, determinisc rails, or transparent agent audits, please share your experience. Let’s co-design behavior specs the way we co-designed REST a decade ago: open, testable, inclusive. The future UI isn’t a screen, it’s a conversation whose guardrails we write together. Human-centered, multilingual, transparently audited. That’s the bar. Let’s clear it, then raise it again.

Comments, critiques, or your own war stories welcome below. Together let’s make sure the next generation of agents feels inevitable on day one, and natural right after.