Why Your AI Agent Is Acting Weird (And How to Fix It)

Most Salesforce professionals are jumping headfirst into Agentforce, but without a clear, structured approach, the AI Agents they build aren’t reliable, often behave unpredictably (or just plain weird), and ultimately never get deployed.

This article introduces Agent Designer, a free tier of the Elements.cloud Change Intelligence Platform that helps you design, build, and govern AI Agents. We’ll break down the importance of scoping AI agents correctly, why instruction design matters, and how Agent Designer can help you design, build, and govern AI agents.

Designing Reliable AI Agents

Use Cases: Complexity vs. Risk

While many conversations around agents focus on appealing, high-profile, customer-facing examples, like the Saks agent – which, for example, helps you return a sweater and swap for a different one – often come with huge deployment risks. 

Such scenarios are typically very complex, involving numerous moving parts and integrations with multiple data sources. And the risk to the reputation of getting it wrong is high.

The smart approach is to pick low-deployment risk agents to start with. It’s important to understand that low deployment risk doesn’t necessarily mean low complexity. For instance, we’ve successfully deployed an internal 26-step, six-action sales coach agent that is complex in its internal workings but carries low deployment risk.

We created an agent that was a “cocktail connoisseur” for a social Salesforce event in NYC. It was relatively low deployment risk because it was just recommending cocktails. It took less than a day to design and deploy, and it was a great showcase for Agentforce. 

Here is a way of thinking about use case selection.

Scope. Scope. Scope.

The scope of the AI agent is defined by the user experience. You could have one AI agent that has numerous Topics that cover a wide range of JTBD (Jobs To Be Done), so the user simply talks to one AI agent. 

Alternatively, you could have an AI agent with a narrower scope and therefore fewer Topics.  But then, the user needs to decide which AI Agent to talk to based on their needs.

The Topic is what delivers a JTBD. Our experience is that Topics are often too broad a scope – e.g., Customer Success. A better Topic scope is “analyze the open case, recommend a solution based on support articles, and write a response for review”. 

If the scope of the Topic is too broad, the AI agent can become confused. The Topic will have a large number of Actions and a huge list of Instructions. This is why AI agents appear to be unreliable – and they are difficult and time-consuming to build, test, and debug. 

Design Matters

Think of your AI agent as a new digital employee. Would you throw a new employee into the wild with no onboarding, no playbook, and vague responsibilities? Would you trust them to do ‘what they think is best’, with no consideration for company practices? Of course not!

Yet, without a clear design, that’s what you are doing when you build an AI agent by adding some actions and a few instructions. The result is an unreliable AI agent, and a huge amount of time wasted during testing, where you keep adding more and more instructions. 

A design-led approach enables you to build, govern, and deploy an AI agent far faster, even though it takes time upfront for the design.

Agents need to be well-architected, with clear consideration for when deterministic actions need to be invoked and when probabilistic reasoning needs to be applied. 

Without clear, well-architected instructions, your AI agent will breach compliance boundaries, struggle with decisions, behave inconsistently, or simply act… weird. That’s not innovation. That’s a risk.

Agent Designer changes the game. It is part of the free tier of Elements.cloud’s Change Intelligence Platform and lets Salesforce professionals design modular, governed, and testable agent instructions all through intuitive diagrams purpose-built for AI behavior. 

Agent Instruction Diagram (AID)

It’s not a flowchart, process map, or a chatbot workflow – it’s a blueprint for a JTBD (Job To Be Done).

Agent Instruction Diagrams let you visually map out, in collaboration with the stakeholders:

• What your AI agent should do.
• How it should behave across different scenarios.
• What deterministic actions should it take and when (your actions).
• What hard rules must it always follow (your guardrails).
• When it needs to perform more complex reasoning, interpretation, or generation (your Prompt Templates).

With Agent Instruction Diagrams, you are mixing prompt engineering with systems design. You’re providing the reasoned flow for an AI agent to operate with judgment, reliability, and compliance.

The Agent Instruction Diagrams automatically generate the Instructions that are used in the AI agent. They help you define the Actions that you need to create for your Agent. We even have embedded AI that validates and suggests improvements to your diagram

To be clear, the AI agent is not following the process like a workflow. It is using the instructions to work out where it is in the process at any point in its conversation so that it can decide how to plan, what response to give, or what action to perform – just like your new employee would.

We believe that you should create the Agent Instruction Diagram even if the AI agent is fairly simple. 

The diagram is more than a design document for the AI agent. It is used to:

• Engage business users and agree scope of the agent and topics.
• Design actions versus instructions versus guardrails.
• Validate and suggest improvements to instructions.
• Generate unambiguous instructions.
• Debug during test/evaluation.
• Evaluate end-to-end scenarios.
• Demonstrate compliance and audit trail.
• Explain behavior, track stakeholder approval, and get signoff.

Proven Approach

This process-led approach is having a dramatic impact on how quickly teams are building agents and getting them deployed. The approach and Elements.cloud are used by Salesforce Professional Services, and they presented on the keynote stage at London TDX. The session was recorded and is available on Salesforce+.  

The approach is now in the Trailhead Agent Planning modules and is being incorporated into the Salesforce Best Practices Guides. We’ve also created detailed step-by-step training.

The notation used in Agent Instruction Diagrams is simple. The magic lies in providing logical, unambiguous, discrete details on both flowlines and activities to inform how the agent should be behaving. Create this by engaging with your stakeholders to design the perfect AI Agent.

Every flowline coming into the activity box serves as a condition. It explains when / under what circumstances an AI Agent should consider performing that particular activity. The activity (Instruction, Action, or Prompt Template) should specify what to do. 

The Time Is Now

The sooner you start building AI agents, the faster you will start building the core skills, creating trust in AI, and identifying high-value use cases for your organization. The great news is that we’ve done a lot of the painful learning, so you can be on the fast track. We’ve got a proven approach that enables you to design, build, and deploy agents quickly and confidently.

And even better, we have made the Elements.cloud Agent Designer capabilities that customers, partners, and Salesforce Professional Services use, free to everyone.

Ready to Try Agent Designer for Free?

Start designing your first AI Agent with free Elements.cloud Agent Designer today. 

Get Started for Free