AWS re:Invent 2025-Building workflows of your choosing: agents, copilots, and everything in between

🦄 Making great presentations more accessible.
This project aims to enhances multilingual accessibility and discoverability while maintaining the integrity of original content. Detailed transcriptions and keyframes preserve the nuances and technical insights that make each session compelling.

Overview

📖 AWS re:Invent 2025-Building workflows of your choosing: agents, copilots, and everything in between

In this video, Brad Rumph, Field CTO at Tines, demonstrates how IT can strategically govern AI-driven workflows by blending deterministic and agentic approaches. He presents findings from Forrester research showing 86% of IT leaders view orchestration as critical for scaling AI. Through a live demo using AWS EC2 and CloudWatch, he shows how Tines integrates AI agents with human-in-the-loop oversight to automatically manage infrastructure—detecting high CPU utilization, creating cases, routing Slack approvals to engineers, and resizing instances from t2.large to t2.2xlarge. The platform runs stateless on AWS Bedrock with Anthropic models, ensuring full auditability while allowing customers to build custom agents using their own prompts and policies without requiring coding expertise.

; This article is entirely auto-generated while preserving the original presentation content as much as possible. Please note that there may be typos or inaccuracies.

Main Part

The Strategic Imperative: Governing AI-Driven Workflows in Modern IT

Morning everyone, and I hope everyone's having a great conference so far. My name is Brad Rumph, and I'm the Field CTO at Tines. It's a bit crazy to think that I've been in this space working with different model-driven tools, iPaaS, API gateways, AIML for the last 25 years, and for the last three years really deep into all things generative AI, GenTech AI, and where we find ourselves with agents today. It's been quite a journey over this time, and now we're working with systems and platforms that are architected radically different. RAG, which is almost a thing of the past, we've got MCP and Model Context Protocol, and even a lot of other protocols that I'm trying to keep up with, like the Agent to Agent Protocol and even the ACP Protocol, which is now being merged into A2A.

So today we're going to cut through all the noise, and what we're really going to focus on is how IT can strategically govern AI-driven workflows. We're going to explore how to blend different types of workflows, from fixed deterministic workflows to highly dynamic agentic ones. And we're also going to touch on where human judgment is added and where that matters the most. So our focus is going to be on infrastructure management using AWS for your infrastructure and then leveraging Tines for your AI orchestration, and how you can securely scale without losing control.

So we all know that AI is moving incredibly fast, and IT is being asked to do more than ever. They are doing things in terms of having to scale systems, managing compliance, and now integrating AI into the mix. And so the opportunity is huge, but so is the complexity, and how do you manage all these things? So the challenge for IT isn't whether to use AI, but how to deploy it safely, strategically, cost-efficiently, in a way that's fully auditable. So agentic workflows, which we define as autonomous systems that are acting across various different tools, really can unlock faster, smarter decision-making while keeping IT in control.

So 86% of IT leaders believe orchestration is critical to scaling AI, and this isn't theoretical. We worked with Forrester, we talked to over 400 folks in the C-suite, SVPs, directors of engineering, all different levels, engineering leaders, and really came up with a bunch of stats, and this was one that really jumped off the page.

And so what does a modern IT environment look like? Well, modern IT environments today are defined by three core pillars, and these all create a very dynamic ecosystem of complexity but also opportunity. If we look at the left over here, and I think they gave me a laser, if we look at infrastructure, this is your digital fabric. This is all of your systems, your data, your APIs, your tools that you want to connect to, from AWS to other specialized SaaS platforms, or maybe you're a hybrid cloud environment, maybe you're running some things on-premises. And we see that a lot with our customers. And this layer generates millions of different signals and alerts and events every day. Well, Tines, we connect to every single one of those, and we can make those events and transform these very raw, high-volume types of alarms and events into highly orchestrated, auditable actions.

If we move over to the right and we look at intelligence, this is what we deem is the cognitive layer. And with the rise of AI, copilots, autonomous agents, these systems can reason and act, but they really need to have some boundaries to operate safely. And so with Tines here, we are the definitive control plane, and we ensure that every intelligent action is governed, explainable, and inherently safe.

If we shift down to the bottom, we look at people. We view this layer and pillar as the strategic operator. And here, these are the decision-makers, these are the folks who are dealing with strategic priorities, having to manage critical exceptions, and really hold the final accountability. With Tines, we embed human judgment into your workflows, and we allow humans to review, approve, and intervene where it makes sense to ensure that the most critical decisions are always informed and timely.

At the intersection, what you see in the middle is Tines. We're an intelligent workflow platform, and we keep IT operations secure, intelligent, and aligned with business goals. This is all wrapped in a layer of control.

Blending Deterministic and Agentic Workflows: Building AI Agents in Tines

If we look at the evolution of workflows, we all rely on highly deterministic workflows. These are fixed logic with predictable results, and these are things for tasks like provisioning or deprovisioning, for example. But what happens when we look at the context and context shifts, or you have to make decisions that depend on nuance or a type of judgment? That's really where agentic workflows can shine, interpreting, reasoning, and adapting. The goal isn't to really replace one with the other, replacing deterministic with agentic. We really think that it's a combination of all of these.

I'd like to introduce this concept of AI agents in Tines. The definition and requirements can vary in terms of what agents look like. You can go to every single booth here and hear somebody talking about what agents have and what they do. The way that we view it is very much like our workflows that you can create in Tines. We believe that these workflows should belong to the people that own them, and those people should be the ones that create them. We're not going to force our own definition of what an AI agent looks like. We're going to give you the tools to build your own agents.

We know that agents should be part of any workflow process, and there's a massive degree of orchestration that has to happen, especially when you roll these things out company-wide. When you do that, prescribed approaches are just not going to work at scale, but the ability to tailor these agents to the individual business needs will. There are a lot of different examples of what this looks like, and we offer pre-built stories out of the box. You can look at those as accelerators, as well as pre-built templates and API integrations.

We allow our customers to take those, tailor their workflows to their needs, and we really provide the same type of ability with building your own agents. Take your standard operating procedures, take your policies, and really focus on building the prompts that differentiate your business from your competitors. We really encourage and work with our customers to help them through that process. The core idea is that you can scale execution without sacrificing any of the oversight.

So how do you decide what type of workflow you're going to use, deterministic or agentic? Well, if risk tolerance is critical, if it has to be perfect every time and highly predictable, then you should stick to deterministic. If flexibility is needed and the outcomes can vary, then agentic may offer some speed and adaptability and really give you that flexibility that you need within your process. Within Tines, we blend both. You can script the rules, you can layer in reasoning, and you can also insert human checkpoints when and where they're necessary.

No automation is complete without human judgment. We often say humans can handle the gray areas, priorities, and exceptions. We believe that human-in-the-loop is not a fail-safe, but it's a deliberate design choice. It's something that we all know, where somebody should really take a look at something. There are those nuances, or maybe the context needs a human to review it. You can choose what runs automatically and what needs review.

We offer a couple of different flavors of agents within Tines, chat-based where a human can be in the loop and interact with that, or it can run completely task-based or autonomously. I think here we're going to tee up a sample use case. I'm going to show you a story that we pulled from our library in Tines, and then we configured it to basically look at some alarms that are going to be taking place, that are monitoring and configured against some of our EC2 instances for our leading FinTech application.

We're really going to focus on when our CPUs are trending or running hot, where there's really high CPU utilization. If the usage and the utilization is normal and we're looking at these patterns,

we'll make certain decisions to maybe just run a flow that operates autonomously. But if it's unusual or abnormal, and something doesn't look right, maybe we're running over a certain CPU percentage or threshold, let's say 96%, well, then we might want a human in the loop. We might want to route a notification to the engineer on call in Slack, right, and have them look at it and decide whether or not they want to approve or disapprove a particular request to upsize or downsize a particular instance. So we look at this as predictive availability, and deterministic workflows can handle routine types of tasks, while humans step in only when something looks off.

So we believe that workflows are evolving, and the future is really a thoughtful combination of human-led, deterministic, and agentic approaches. So we don't put you in one box or the other. We give you the full autonomy to choose what makes sense so you can blend all three together within a single workflow or automation. And we believe that intelligent workflows are truly the key to unlocking AI's true potential.

So before we hop into this story that I mentioned, and again a story is a workflow in Tines speak, it's worthwhile to note that Tines builds all our AI features with trust front and center. All our AI runs stateless, private, in region, it's tenant scope, there's no networking, training, storage, or logging of any of your data. And you can use our Tines hosted model. We partner with Anthropic, we run on top of AWS Bedrock, or you can bring your own models from OpenAI or Google Gemini, Mistral, Llama. So we give you a lot of options, but again, out of the box, we provide Anthropic models under the hood.

Live Demo: Intelligent Infrastructure Management with Human-in-the-Loop Oversight

So your fearless leader, Gavin Belson, he's becoming incredibly worried lately that Pied Piper is figuring out how to use intelligent workflows and AI agents to really scale and manage their infrastructure much more efficiently than Hooli's. And so he's tasked his engineering team and SRE teams to look into how AI agents can be inserted into workflows and play a much bigger role in helping to manage and maintain Hooli's infrastructure. So what you're going to see is intelligent workflow automation and orchestration with the human in the loop, AI agents, and cases within Tines.

So we're going to start off by looking at our Hooli PROD-1 instance in EC2. We can see that it's running. Instance type is a t2.large. We're going to pop over into Tines. This is a Tines story, right? So at the top, we're looking for alarms that we're getting from AWS CloudWatch. And what we have is a couple of different triggers, basically conditionals, go this way or that way, where on the left-hand side, if CPU utilization is greater than 90% but less than 96%, we're going to go down that path. Or if it's greater than 96%, we're going to go down the path on the right-hand side over here.

If we go down the right, we're going to create a case. We're going to send it for approval in Slack. If it's approved, we're going to update the case, and we're going to update the instance size by upsizing or downsizing it in EC2. And up here you can see that event just came in from CloudWatch. So an alarm, we get this alarm. We see that in that alarm, when we look at it, it's greater than 96% CPU utilization. We're going to create a case in Tines. So we're using an AI agent to do that. We have some system instructions. We have a very simple prompt that, hey, all we want to do is create a case in Tines, and we give it some metadata.

When I come over and I look at what the agent is actually doing, we can look and see exactly the decisions that it's making at runtime. So we can see where it's created a case, it's adding all the metadata for that particular EC2 instance. And when we pop back over into the story, we can see that the case update's been completed, and we've now progressed to sending an approval to the engineer on call in Slack. He can either choose to approve or not approve this request.

So we're going to pop over to Slack. You can see we've got this channel set up, notification came in, we have an abnormal EC2 instance, over 96% instance name, instance ID, and we have a link to the case details that our AI agent created where you can get additional details. We ask the engineer if he wants to approve, he clicks yes. Come back into the story, you can now see that we've progressed here to, hey, determine what the instance size should be and update that particular instance accordingly. So again, we can look at all the events that are flowing

through these stories. Again, another very simple prompt in terms of what we're instructing the agent to do. Here I was lazy and actually hardcoded the instance ID, but earlier I pulled it in through some metadata. We can look here and see what the instance details are. We can see that it's currently a type t2.large, as we saw in the very beginning of the presentation. What VPC, what subnets. And here we can get the scaling decision and implementation that the agent, based on our prompt, is looking at. Here we get a summary of the actions taken. And you can see that the agent has made the decision to upsize the particular instance from a t2.large to a t2.2xlarge.

We come back to the story, and we're now going to update the Tines case. Again, we can see exactly what's happening while this is running. We scroll down to the bottom, and we can see that the agent's thinking. It's going to give a case update summary. It says what its resolution information is, what the root cause was of why it made the decision that it did. And I think it's very important to note that all of these events, we store them. So you can retrieve them at any given time, hence the auditability capabilities and the governance that we provide within the platform.

This case has been done. It's added some tags automatically. The agent came up with these just based on what the events look like. Here you can kind of see just kind of a recap of that particular case. High CPU utilization was detected. Here's what the previous state was, the configuration, here's the resolution and the recommendations that were made for the upgrade. We'll go back over to EC2, we'll look at our HOOLI-PROD-1 instance, and we can see that it's now running and it's been upgraded to a t2.2xlarge.

So what did you just see? What you just saw is how you can seamlessly build deterministic and agentic workflows with human-in-the-loop and human oversight added in. I showed you how easy it was to configure an AI agent within Tines. I think it took me literally a couple of minutes to write the prompt. It needed some tweaking, but we integrated directly with Amazon's Claude to do all the upsizing that you saw. And then in terms of cases, we have really robust case management capabilities within the Tines platform, and that really enables all those on-call engineers or anybody within the organization to collaborate on that particular case and see exactly what happened. And with very little instructions within the prompt, we got a lot of detail auto-populated automatically within that case in Tines.

So with Tines acting as the intelligent workflow platform, even Gilfoyle and Dinesh can agree that you don't need to be an expert in Java, Python, or scripting for that matter, to eliminate muck work within your organization. And with Tines, Gilfoyle knows it's time to truly move on from Anton and make the move to AWS. So this isn't about launching a single AI agent. It's about governing every workflow from end to end. IT shifts from being a gatekeeper to an orchestrator, secure, intelligent execution at scale. And so with AWS and Tines and your policies, your SOPs, and the prompts that you create, you can gain speed without losing control. AWS provides the reliability where Tines really provides the control. Your policies and prompts define how it all comes together in the end.

So I'm going to leave you with this question, and I want you to think about it. From an infrastructure perspective, what's one process or more, one or more processes that you would leave to run completely ungoverned and completely autonomously? And what are some processes that you would actually deem a human should be in the loop to provide that oversight? And that's really where orchestration meets opportunity. So we say the best way to understand Tines is to see Tines. So I want to encourage everyone to come by booth 1849. I'd be happy to carry the conversation on or answer any questions that you have. Appreciate you all coming out, and I hope you enjoy the rest of the conference. Thank you.

---

; This article is entirely auto-generated using Amazon Bedrock.