In December I wrote down three predictions for 2026: multi-agent orchestration, enabling AI company-wide, and everyone building their own agents. Half the year in, here’s where each one actually stands.
Orchestration is exhausting
The shift from single agents to multiple agents working in parallel is already here. Many engineers have figured it out and they’re more productive because of it.
What you only discover later is the fatigue. A senior engineer last week described it as “running four projects at the same time and context-switching between them every minute.” That’s not a tooling problem. It’s what coordinating parallel agents actually feels like in practice.
The next thing we’ll learn, I think, is how to tell an agent when it’s done. Not just “the tests pass” but “the work matches what was asked, the verification ran, the guardrails held, stop.” Right now most of us are doing that verification ourselves. The whole process becomes less stressful when the agent learns to continue until it can verify and close it too.
Enable the company, or lose the people
This is the one I was too optimistic on. I expected the bottleneck to be vendor selection. It’s not. The bottleneck is that most enterprise AI tools aren’t enterprise ready, and the gap between what your people can do with AI at home and what IT will let them do at work is widening every week.
That’s the talent risk nobody’s putting on slides. The best people aren’t waiting for IT to catch up. They’re updating their CVs and asking whether the next company has figured this out yet. The window for closing the gap safely is shorter than most CTOs think.
What is different in six months
The AI technology moved as fast as I expected but the enterprise tooling and the organisation moved slower, and the gap between the two is where the actual challenge is now. None of the three bets are wrong. They’re just harder, slower, and more dependent on each other than the December version of me was willing to admit.
