Stop doing this, I’m frustrated with y’all#
Have you ever been asked, “What’s a topic you could give a TED talk on with zero preparation?” This blog is one of mine: why some agentic implementations drown and others make Michael Phelps look like he needs floaties.
What made me feel the need to rant about this is that I see both patterns in my work with customers at Temporal. There is no shortage of stories and metrics publicly shared about struggling implementations. Those stories match what I see from teams trying to build without Temporal as their harness. I see those stories, nod sadly, then move on to helping my customers who are actually live and building agents that transform their business.
The real destination is to build an Intelligent Business: one that gets the right data to the right decision-maker, so they can make the right decisions quickly. Sometimes the decision-maker is a human. Sometimes it is an ambient agent. Sometimes it is humans working in concert with agents. It does not have to be an internal decision-maker either. Your customers may benefit from making decisions about how to use your products and services with help from a well-informed agentic guide, especially for customer service, evaluation, and usage guidance.
The alternative to the Intelligent Business is data silos, uninformed or misinformed decision-makers, slow actions, lack of agility, and a persistent feeling of being underinformed.
Intelligent Business: one that gets the right data to the right decision-maker, so they can make the right decisions quickly.
I’ve experienced both the lack of good data and the impact of having the data I need. It can look like “we don’t know what to do because we don’t know what is actually happening,” “I know the data is over there but I can’t read it all,” or “I just don’t have access to what I need to make the decision.”
The empowered, intelligent case feels like this: “I am making an informed decision because I understand the data and I have access to what I need. I understand the shape of this information and can drill down to validate specific areas.” It feels good to be more informed than you ever were before, and this is completely possible in this agentic era.
Agents are most powerful when they do what they do best: process huge volumes of data, summarize and explain it to humans who can then understand, decide, and act. Ambient agents are also powerful, but they must explain their reasoning and be visible, auditable, and understandable.
One common “we went too far” anti-pattern I see is delegating understanding to agents. Humans are completely out of the loop, out of touch, with no ability to know what agents are doing. This leads to employees saying things like:
“What I did doesn’t matter because the agents are faster and smarter than me and I don’t understand what they did. When they succeed I don’t know why and I am not a part of that success, and when they fail, I feel like I’ve failed but I don’t know why.”
So you’re building an AI platform team, or you should be#
This post is for people who are trying to get to the state where their business is taking full advantage of agentic capabilities. Maybe you are on an agentic engineering team. Maybe you work in data or software engineering and are curious. Maybe you have just been tasked with building all the agents your company needs, or you are part of the founding membership of an AI platform team.
This work is powerful and important, and I strongly encourage large organizations to build AI platform teams. This is worth making a durable capability of your business. Establishing agentic engineering practices, enabling the large numbers of agents you should be building, and setting governance and guidance are key tasks.
Things you need to build#
In my experience, agent architectures need the following to be successful and easy to build:
- Simple-to-use memory and context
- Built-in resilience to failure
- A simple programming model with strong built-ins for events, queryability, and interaction
- A simple-to-understand development model built as code, not graphs
- I personally hate graphs, but if you love them, I won’t judge you; have at it.
- Keep in mind that coding agents are quite good at writing code.
- Token efficiency
- Auditability for every agent interaction and decision
- Sampling is not enough.
- This is especially true for customer-facing agents.
- Visibility into every running agentic process
- Useful for both debugging and customer-facing agents.
- The ability to run for an unlimited time
- Without losing memory or context.
- Strong data typing
- Durable LLM and tool orchestration
- Simple customization, ideally with customization hooks
- Full orchestration control: own the code to your harness, and don’t get locked in to a specific harness, a provider’s models, or required architecture changes
- For example, this lets you build your own guardrails as needed.
These needs map naturally to Temporal’s core primitives: Workflow Executions for durable orchestration, Activities for external calls and tools, and Signals, Queries, and Updates for interaction with running processes.
The lack of these things causes great struggle. Often, simple things like LLM rate limiting, API failures, and unpredictable agent and human behaviors become core causes of production failures.
Things you need to do#
It is not enough to build the harness. AI platform teams also have to grow adoption, trust, and organizational momentum.
How to succeed at agentic platforms#
The primary importance of secured data#
The first thing successful agentic implementers do is get their data security right. Data is loaded via tools that should run with user-level authentication and authorization. Data providers should respect user-level access. You don’t want to use service-level authentication here. Maybe your data API has access to a lot of data you don’t want every user to have.
For example, for an agent that helps your customers manage their own orders, you wouldn’t want to expose all orders. You would want to restrict the agent to their data.
I recommend using OAuth and OAuth-aware MCP servers. Building this access layer is a significant effort, but it is required before agents can be successful and secure.
A lack of user-level secured data means you get one of two outcomes:
- Agents that can only use publicly or generally available data
- Agents that can expose the wrong data to the wrong people
Deployed agents vs. desktop agents and why it matters#
On the subject of data: desktop-based agents have been powerful general-purpose tools because they can access local data, such as code and documents. But they often have a limitation of directed context: they are general purpose.
For specific business workflows, my recommendation is to have a more opinionated agent deployed in your infrastructure, with your guardrails and your security profiles. We call these “deployed agents.” Deployed agents need a harness and, wait for it…
Temporal Applications make for a great harness#
…Temporal is a great harness foundation.
Temporal gives you Durable Execution, memory, an easy-to-use programming model, unlimited runtime, testability, and the ability to define everything as code. Coding agents are great at writing high-quality code.
Temporal also gives you the building blocks to run and evolve production systems in your own environment, including Worker deployments, Event History, and the Temporal AI Cookbook for practical agent patterns.
Using Temporal as the core of your agentic platform#
Agents are most powerful when they do what they do best: process huge volumes of data, summarize and explain it to humans who can then understand, decide, and act. If you want to start building agents that transform your business, and you want to focus on that transformation instead of struggling with harness problems, we have resources for you.
Here is a pre-built harness you can customize to get the value of Temporal quickly, with hooks you can implement for guardrails, tools, and more. Check out this reference sample: https://github.com/mikeacjones/temporal-agent-harness-example/tree/main
Check out the AI Engineering section of our reference platform hub: https://go.temporal.io/platform-hub/ai-engineering
For more context on agent architecture and production patterns, see Making friends with agents: A mental model for Agentic AI applications, Durable Execution meets AI: Why Temporal is the perfect foundation for AI agent and generative AI applications, and Of course you can build dynamic AI agents with Temporal.
You don’t have to take my word for it. Here Cat Zhang from Block explains how they use Temporal to transform Block with AI agents:
Block @ Replay 2026 | Powering AI Agents with Temporal and Kgoose
I wish you great success in the transformation of your company into an Intelligent Business.