THE AGENTIC FUTURE WITH TEPORAL: A FIRESIDE CHAT 

Last week, we hosted a tech meetup at Essent's offices in Den Bosch, exploring the relevance of today's software engineering wisdom in the agentic future. We brought together developers, architects, and technology leaders to discuss the practical realities of implementing agentic AI systems and what this shift means for software engineering.

In this post, I'd like to capture the key takeaways from the fireside chat about Agentic AI, and the path ahead of us. This is a joint blog post between Temporal, Essent, and Navara.

We're in the "wild west" of agentic AI

One of the most honest takeaways from our discussion acknowledged that we're in the "Wild West" phase of agentic systems. This field moves fast, with constant uncertainty and few established best practices.

For both engineers and decision makers, this means adaptability and willingness to experiment are vital. There's no playbook yet, and what works today might be outdated in six months. This uncertainty isn't necessarily bad news. It creates opportunities for organizations willing to invest in Agents and systematic experimentation.

Testing when things aren't predictable or repeatable

One of our liveliest discussions was about quality assurance for agentic systems. Traditional software testing breaks down when the same input doesn't always give the same output. Josh (Engineering Manager - Conversational AI, Sinch) talked about how QA for their agentic systems rely heavily on "evals" (evaluations) and "guardrails," which are often implemented using LLMs themselves. This creates interesting new challenges around cost, reliability, and defining what "good enough" means. When you can always improve aspects by spending more tokens, when do you stop? When does certainty reach a level that is acceptable?

We all agreed on the importance of making a conscious choice about what kind of task you delegate to an LLM, and which tasks you should delegate to deterministic and tested code. By lowering the set of actions an agent can take, and by implementing deterministic guardrails in which that action can be executed, it's possible to control the impact of how far an agent is able to deviate from the intent we want it to deliver on. This approach resembles the way we treat humans a lot: It's a very bad idea to give every human agent raw direct administrative access to your production database, so why should we trust an agent with it? Rather, we should focus our efforts on making sure the agent is able to request data mutations or queries in the exact way that makes sense in the context of achieving its goal. This adds additional demands to our security architecture and internal documentation, but those demands do not differ from anything we consider to be good engineering practices today.

The bottom line: We don't have established best practices for testing the LLM components of agentic systems yet. The deterministic parts, especially the actions taken by the agent, can still be tested with

traditional methods. We're all figuring out as we go how to handle quality assurance for the unpredictable, AI-driven parts, and we should limit the extent to which it can go off-track as a result.

Most enterprises struggle with AI agent implementation

A major theme in our conversations was this infrastructure readiness gap. Recent McKinsey research indicates that while 78% of companies are using generative AI, over 80% report no material earnings impact¹, and fewer than 10% of vertical AI use cases make it past the pilot stage¹. This suggests that the AI technology itself isn't the bottleneck: It's the underlying IT foundations that aren't up to the task. To deliver a high-quality agent that does exactly what it's supposed to, we need to provide it with the right tools and context to complete its goal. We also should not have to completely replace existing software with AI-driven replacement but rather augment them.

The usual suspects are to blame:

● Legacy workflow tools that can't be augmented with dynamic AI-driven orchestration

● Poorly documented interfaces that break reliable system integration

● Manual edge-case handling that stops automated workflows dead and loops in humans

● Weak internal security that creates vulnerabilities

The result: Based on early implementation experiences and industry data, organizations see 3-5x longer implementation times and frequent project failures when they try to deploy AI agents without proper groundwork. McKinsey reports that many companies are even "retrenching — rehiring people where agents have failed."²

Dynamic vs. static workflows

We had a great discussion about the trade-offs between different workflow approaches:

● Dynamic workflows are flexible but harder to predict and test

● Static workflows are more reliable but less adaptable

The consensus: A hybrid approach probably works best. Start with dynamic, agentic workflows to explore what's possible, then lock in successful patterns as static workflows once you understand them better. The reverse approach can be applied to existing workflows. Enable agents to assist or replace human interventions, and lock in the successful patterns as well.

This fits in extremely well with Temporal's platform, which supports both flexibility and reliability without forcing you to choose one or the other³.

From prompt engineering to context engineering

Our conversation highlighted an important shift in how we work with LLMs. The field is moving from "prompt engineering" to "context engineering," focusing on giving LLMs exactly the right amount of context to get better results and reduce hallucinations.

This represents a real shift toward treating context as a strategic resource that we can build up from previous executions and event logs to make future agent behavior better.

The learn-codify-optimize cycle

But how do you provide the right context and actions to an agent? Especially if you're designing future "AI-Ready" systems:

1. Deploy agents for new processes (or existing process augmentations)

2. Capture decision patterns and outcomes

3. Analyze what works and what doesn't

4. Codify successful patterns into deterministic workflows

5. Scale across the organization

This is what organizations are doing to gradually build competitive advantages through adaptive, self-improving business processes. It's also a cycle that can be heavily optimized by letting several agents collaborate. A process-oriented AI agent can reduce manual labor inside processes, while an analytical learning agent can help to capture successful patterns. A coding agent can then suggest (deterministic) process improvements to the codebase, so the first agent is no longer required to step in. If this process turns into a well-oiled machine, it starts to provide agility and iteration cycles that will be very hard to beat as a human.

Use a foundation-first approach

In all these scenarios, there is one constant: a need for increasingly solid software engineering practices. Multiple speakers stressed that organizations need proper foundations embedded in their action library before trying to implement agentic systems. It's a very dangerous risk when companies start to consider agents as a full replacement for software engineering instead of an augmentation.

The insight: Poorly designed legacy systems don't just slow you down: they create serious risks when autonomous agents start accessing them at scale.

Good engineering fundamentals remain critical, even as new paradigms emerge. Organizations shouldn't abandon best practices while chasing innovation.

The economics matter

When you start to run LLMs for continuous evaluation and guardrails, it quickly gets expensive, and organizations need to factor that into their planning. We need to start treating tokens as a resource. How many tokens do we want to spend on security? How many on testing? And how many on the business process itself? Do we want to spend tokens on creating a deterministic workflow, or do we want to spend tokens on being inside the process?

The entire group talked about finding a justification for spending LLM-credits: weighing flexibility against computational costs and unpredictability risks. This economic reality pushes us to systematically convert successful agentic behaviors into deterministic code once we understand the patterns.

The window is closing

Throughout the event, we all emphasized the urgency. Industry analysis suggests that leading organizations started their foundation investments 12-18 months ago, and every month of delay increases transformation costs exponentially. McKinsey notes that fewer than 30% of companies have

CEO-sponsored AI agendas¹, indicating that many organizations are still in the early stages of strategic AI adoption.

We're at a critical point where organizations face a clear choice: invest in proper foundations now for competitive advantage, or delay and permanently fall behind competition.

What's next: practical steps

For organizations ready to start their agentic AI journey, our discussions made the path forward clear:

1. Start with infrastructure assessment - Understand your current capabilities and gaps

2. Invest in API-first architecture - Make sure all systems can be accessed reliably by both humans and agents

3. Implement dynamic orchestration - Choose platforms that support both deterministic and intelligent components inside the same workflows

4. Begin with human augmentation - Enhance existing decision points rather than replacing entire processes

5. Establish learning cycles - Create systematic approaches to capture and codify successful patterns

Keep the conversation going

Agentic AI is moving fast, and no single organization has all the answers. We need to keep sharing knowledge, learning from each other's experiences, and collectively figuring out what works.

The future belongs to organizations that can combine human judgment with intelligent automation in ways that adapt and improve over time. The infrastructure decisions you make today will determine your competitive position for the next decade.

The question isn't whether agentic AI will transform your industry: it's whether you'll lead that transformation or struggle to catch up.

References and notes

¹ McKinsey & Company. "Seizing the agentic AI advantage." June 13, 2025. 

² McKinsey & Company. "One year of agentic AI: Six lessons from the people doing the work." September 12, 2025. 

³ Temporal Technologies. "The fallacy of the graph: Why your next workflow should be code, not a diagram." August 20, 2025.

The field of agentic AI is rapidly evolving, with most implementation data coming from early pilot programs and organizational case studies rather than comprehensive peer-reviewed research. Statistical claims reflect industry observations and early adopter experiences as of September 2025.