Agentic AI in 2026

2025 was all about agents and gave us a glimpse of what is possible with agentic AI. It’s easy to forget though that we are still in the early days. There is no doubt that adoption will only accelerate in 2026 across all industries and company sizes.

Looking ahead, we asked ourselves a simple question:

What are the main challenges that need to be solved so any company can successfully roll out a fleet of AI agents working around the clock?

These challenges form the basis for our top predictions for agentic AI in 2026.

1. Agentic search becomes increasingly important

A lot of knowledge work involves retrieving the right information scattered across inboxes, files, databases or the web, followed by applying business logic on top of the relevant information. Providing the right context is therefore a prerequisite for the agent to do meaningful work.

To find the correct information, we need a holistic search system operating at scale across data types (from spreadsheets to video), and with the ability to evaluate and refine search results. Most popular AI automation solutions still fail spectacularly when it comes to search, limiting their usefulness for real work.

As detailed in a previous blog post, we see agentic search as a core building block for agentic AI.

2. Durable execution for long-running tasks becomes critical

As agents become deeply embedded in companies, they become part of long-running tasks which are interleaved with other business processes and have humans in the loop.

This means that an agent might realize it needs to request extra information or wait for the result from another process while it’s in the middle of doing a task. Once this missing information is available, the agent picks up right where it left off. This type of start-stop behaviour could happen frequently and means that the agent possibly remains active for long periods of time which requires durable execution. Anthropic for instance mentioned that they already observed Claude Sonnet 4.5 maintaining focus for more than 30 hours on coding tasks.

In the coming weeks, we will detail how we solve some of these infrastructure challenges internally.

3. Workflows and agents will be united

Workflows vs. agents is a topic of discussion that comes up every so often. This discussion is however misguided and ignores that there is a continuous spectrum between a fully deterministic sequence of steps in a workflow and open-ended agents where the agent needs to decide on which tools to use and in which order for every single instance. Almost all real work sits somewhere in the middle. Tasks can typically be broken down into steps which are deterministic and those that are more open-ended. AI agents need to reflect this reality and combine the best of both worlds. This drastically improves efficiency, speed and reliability because fewer tokens need to be generated, more tokens can be cached, and unnecessary LLM prediction variance is eliminated.

Check out Why workflows are here to stay for a deep dive on this topic.

4. Domain experts will be managing teams of agents

Continuous evaluation has always been key to successfully deploying AI systems. This is no different for AI agents. We do however anticipate that the responsibility of evaluating and improving the agents’ performance will increasingly shift from developers or data scientists to domain experts. There are a few main reasons for this:

1. The output produced by agents looks similar to work done by a domain expert and is therefore best judged and improved by the expert, not the developer.
2. The skill of training bespoke machine learning models has largely been eliminated for many tasks and replaced with simply defining the task logic. Domain experts are better positioned to do this than developers.

This shift means that tools for managing and improving agents need to be tailored to domain experts and business people. These tools need to identify and explain how changes in the data and behaviour of AI agents impacts business metrics and outcomes. This is the only way to systematically improve agents at scale. One example of this is our insurance portfolio root-cause analysis agent which automatically figures out why portfolio metrics are changing, allowing underwriting and claims teams to take the right action straight away.

---

There are of course many other open challenges such as multi-agent orchestration, parallel task execution or optimizing the UX for people to interact with agents. At Decision Computing we are working on many of these topics, so reach out if you want to discuss more and let us know what you think 2026 has in store for AI agents!

Happy New Year!

The Decision Computing team