Five AI trends we're watching for the rest of 2026

Every week brings a new model, a new benchmark and a new headline. Most of it is noise. A few shifts, though, are changing how software actually gets built and run, and they're the ones shaping our roadmap. Here are five we're paying close attention to.

1. Agents are moving from chat to work

The first wave of AI tools answered questions. The current wave finishes tasks: it reads the ticket, opens the files, runs the tests and comes back with a change to review. The interesting question is no longer "can the model write this?" but "can the system around the model check its own work?"

2. One model is giving way to many

Instead of a single assistant doing everything, more teams are orchestrating groups of specialised agents: one plans, others build, another reviews. It mirrors how good engineering teams already work, and it's the idea behind the agentic swarms we use across our own products.

3. Evaluation is the new unit test

When part of your system is probabilistic, "it worked when I tried it" isn't good enough. The teams pulling ahead treat evaluations like tests: written up front, run on every change and tracked over time. If you can't measure whether an AI feature got better, you can't safely ship it.

4. Smaller, faster models where it counts

Not every job needs the biggest model. Routing simple, high-volume work to smaller models keeps responses fast and costs predictable, and saves the heavyweight models for the problems that genuinely need them.

5. Governance is becoming a feature

Customers increasingly ask how an AI feature was built, what data it touches and who is accountable when it's wrong. Standards such as ISO/IEC 42001 give those answers a shared shape. We see responsible AI less as paperwork and more as something buyers will expect to see.

What it means for us

• We design features around agents that do the work, with people reviewing the result.
• We write evaluations alongside the feature, not after it ships.
• We pick the smallest model that does the job well.
• We treat security and AI governance as part of the product, not an afterthought.