Agents will need to get serious about teamwork in 2026

Three years since ChatGPT supercharged the AI field, predicting the future of the fast-changing technology is not for the faint of heart. But one sure bet is that, in the last weeks of the year, our inboxes will bulge with experts who want to give it a go anyway.

Many of last year’s predictions concerned the onset of AI agents. Now they’re more or less here, and businesses will be increasingly focused on getting them to work together, trusting them, and dealing with the risks they might present, our experts seem to agree.

There remains an undercurrent of worry about ROI still being somewhat TBD, and fears of an AI bubble have heightened in recent months. World models, another not-entirely-new term that’s somewhat definitionally vague, are suddenly seeing more interest as a potential way to give LLMs better bearings.

Here are some of the major themes that tech pros told us might define AI development in 2026.

Agents in concert

Those AI agents that companies have “hired” may be in need of some trust falls at a corporate retreat. Coordinating between agents and making them more trustworthy were big topics among prognosticators.

How businesses are able to array agents to handle complex processes will be more important than tapping the most sophisticated model, PwC’s advisory chief technology and information officer, Vikas Agarwal, told us.

“People that look at that whole chain…will have a lot of measurable outcomes,” Agarwal said. “And I believe there are going to be winners and losers in a separation of people that think about the approaches from that way.”

This approach will hinge on coordinated efforts between various companies, including fleshing out nascent communication protocols like MCP and A2A that help agents talk to each other and with various business tools. In December, OpenAI, Block, and Anthropic formed an org called the Agentic AI Foundation to establish these sorts of open standards and oversee MCP.

“The whole point is to take the repeated tasks that are toil and elevate the human to higher-level outcomes,” Arnab Bose, chief product officer at Asana, told us. “You won’t achieve that unless these agents start working together, and in theory, the same kind of protocols and standards that emerged in the first wave of enterprises being built on cloud-native software—you’d end up with similar protocols.”

Trust concerns

But roadblocks remain in even getting businesses to the point where they might let agents handle core business functions; questions abound about errors in their work, security, and general trustworthiness.

“I’m both excited and worried about the maturation of agents,” Peter Lenke, a managing partner at Microsoft’s M12 venture fund, said in an email. “[It] feels like we are still in early innings—there is a lot of promise, but there’s not yet widespread adoption, especially at the enterprise [level]. I think there will need to be much more education on the tooling, organization, and activation of these new agentic tools for enterprise workers.”

Model race slowdown?

Experts predicted a move away from racing toward the biggest and most sophisticated model. Businesses will instead look to eke out improvements from a mix of existing models and agents.

“The era of simply throwing more compute and data at pre-training to build ever-larger foundation models is ending,” Vivek Raghunathan, SVP of engineering at Snowflake, said in an email. “We’re running out of high-quality pre-training data, and the token horizons needed for training are becoming unmanageably long…The focus in 2026 won’t be on sheer size of AI models, but on refining and specializing models with techniques like reinforcement learning to make them dramatically more capable for specific tasks.”

Zoom CTO Xuedong Huang said that companies need to assemble a mix of different models that perform best at different tasks in the most efficient way.

“This is really a complex landscape,” Huang told us. “The ability to federate, to bring the best of the best together with different models—open-source and small language models—you have to really adjust dynamically.”

Bubble watch

At the same time, tech giants like OpenAI, Google, Meta, and Microsoft have poured eye-popping sums into new AI infrastructure in the back half of this year. Investors remain on alert for signs that this spending might be a bubble ready to burst.

Anthony Enzor-Demeo, newly appointed CEO of the Mozilla Corporation, said a bubble may also pop as AI companies try to charge more subscription fees to recoup costs and publishers lock down content to protect it from scrapers.

“This shift is creating a two-tiered internet, much like the difference between flying first-class with a personal concierge and being restricted to navigating public transportation with a frequently delayed, outdated map,” he said in an email. “This could become a bubble bigger than dot-com, potentially bursting if high-cost solutions fail to achieve widespread adoption and revenue needed to sustain their current valuations.”