Where agents stand today
The module "What is an 'AI agent,' really?" showed what a single agent can do today: an AI that uses tools independently and handles multi-step but narrowly-scoped tasks (see also "When multiple AI agents work together" for teams of agents). The current research focus goes further: how far can an agent plan and act on its own, without needing a human at every step?
The research leap: long-horizon planning and tool use
The active research direction is called "long-horizon agentic AI" - systems that break a complex task into many intermediate steps, choose different tools for themselves (databases, software, other AI systems), and keep working consistently over hours or days, instead of just answering a single request. Early research prototypes already show this on narrowly-scoped, well-verifiable tasks - coding tasks whose result can be tested automatically, for example.
The real limit: errors compound
The central unsolved problem isn't any single capability, but the chain: every intermediate step carries a small error probability. Over ten or twenty steps, that adds up - an agent that's 95% correct at every step comes out completely error-free in only about a third of cases after 20 steps. That's exactly why real verification (see "How to really check if an AI solution works") becomes a core prerequisite, not an afterthought.
A realistic take: when will this become everyday-usable?
For narrowly-scoped, automatically verifiable tasks (e.g. code, rule-based data entry), first production deployments are already underway today. For open-ended, business-critical multi-step tasks without automatic success verification, broad, reliable maturity is realistically still several years out - contingent on whether the verification problem gets fundamentally solved. Anyone promising a "fully autonomous AI worker for everything" today is skipping right past this limitation.
Why this matters for you as a decision-maker
The more autonomous agents become, the more important the surrounding infrastructure gets: clear approval levels (see "Human in the loop: when AI needs approval"), logging, and a defined process for when something goes wrong (see "Security in production use"). That's not a nice-to-have for "later" - it's the precondition for actually benefiting from what's coming, instead of being caught off guard by it.