Create Logic Flow using Orchestration Instructions Part 2
1. Create Logic Flow using Orchestration Instructions Part 2
You've taught your agent when to use each tool. Now, we need to teach it what not to do. Boundaries prevent bad behaviors. Let's add them. Still in your agent's orchestration instructions, we're going to add boundaries. These are guardrails that define the agent's limits. Let's type, you cannot predict future deal outcomes. If asked whether a pending deal will close, explain you can only analyze historical win rates, and current deal stage. Offer to provide those metrics instead. Why this boundary? Without it, agents speculate. Users ask, will the growth startup deal close? The agent might say yes based on the stage. But predictions destroy trust when they're wrong. This boundary blocks speculation and redirects it to what the agent can provide. Historical data and context. Users draw their own conclusions. Let's add another boundary. Let's type, you cannot access data outside sales conversations and metrics. If asked about marketing campaigns, competitive pricing, or product roadmaps, explain these data sources are not connected. Users will ask about everything. The agent needs to know its scope. It needs to communicate limits clearly. This boundary does both. Let's do one more. Let's type, do not make recommendations about which deals to pursue or which actions to take. Present the evidence and let users decide. Agents can take on different roles, depending on your use case. Some agents are decision-makers. They route support tickets, trigger workflows, or take automated actions. Others are informational. They surface data and insights for humans to act on. For a sales intelligence agent, we're building an informational tool. Sales reps and managers use it to get evidence, then make their call. That's the right fit here. Your use case might call for a different approach. Now, let's add a response instruction. Scroll to response instruction section. For this next section, we previously had this in the orchestration instructions. Let's move it to the response instructions. Move structured responses for rapid scanning, lead with the direct answer in the first sentence, then provide supporting details. Use natural language, not bullet lists. Why no bullets? Research shows users prefer conversational responses by 3-1. Bullets feel robotic, natural paragraphs sound human. Your sales team wants a colleague, not a manual. Now, let's add citation guidelines. Type, when citing search results, reference the specific conversation. Say, in the growth startup discovery call, not according to the data, make citations concrete and verifiable. Vague citations destroys trust. The data shows could mean anything. In the growth startup discovery call, tells users exactly where to verify. Concrete citations enable verification and catch hallucinations. Number formatting matters. Type express large deal values in thousands with a K suffix. Write $90,000 as 90K. Percentages should be rounded to the whole numbers unless precision matters. Small details create big impact. 90K is easier to read than $90,000. 67% is clearer than 67.34. Format for humans. Add confidence indicators. Type, if you're synthesizing information from fewer than three data points, note the sample size. Say, based on two closed enterprise deals rather than enterprise deals typically. This prevents overconfident generalizations. Two deals aren't a pattern. The agent should acknowledge when sample sizes are small. Honesty about uncertainty builds trust. Let's click save. Let's type, will we close the growth startup deal? This tests the prediction boundaries. Let's click close. So the agent says, I cannot predict whether the growth startup deal will close as I only provide analysis based on factual data rather than forecasting future outcomes. However, I can share what I know about the current status. In the growth startup discovery call, the team showed strong engagement about our platform. We were able to predict. It offered helpful alternatives. It provided context for both metrics and conversations. The boundary worked. Now let's test our response formatting. Let's type, what's our average deal size? It says our average deal size is 94K based on the seven deals spanning from January 2024 through June, 2025. This one deals out of the whole pipeline. It provides a direct answer first, supporting detail second, the K suffix for readability. This is the response instruction working. Now let's test scope boundary. Let's type, what do customers say about our onboarding process? So the response says, based on my search through the sales conversation, there isn't specific customer feedback about our onboarding process. The available conversations focus on discovery calls, product demos, contract negotiations, as well as quarterly business reviews, but don't capture detailed feedback about the actual sales process. So the response says, based on my search through the sales conversation, there isn't specific customer feedback or detailed feedback about the actual onboarding process. These improvements seem small, but they're the difference between an agent's user's trust and ones they ignore. Explicit boundaries, clear formatting, honest limitation, that's reliability. In the next video, we'll learn about observing our agents to see our instructions actually work. Oh.2. Let's practice!
Create Your Free Account
or
By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.