AGI in 2026: how to interpret advances in AI agents

What task-horizon tests measure

Discussion of AGI often compresses many capabilities into one question: “Are we there yet?” Research and engineering start with the task: what must the system do, what information does it receive, and how do we know it finished correctly?

METR tracks a task-completion time horizon: the estimated time a human expert would need for a task on which an agent reaches a specified success probability, such as 50% or 80%. This is not the time the agent actually spent running. The tasks focus mainly on software, machine learning and cybersecurity. METR also cautions that estimates above 16 hours are still unreliable with its current task set. Source: METR [1].

The “hours” measure therefore has to be read together with its methodology. Otherwise an experimental result can become a commercial promise the test never supported.

Our reading: three questions before extrapolating

At Polybee, we separate three questions when interpreting an advance. This is an editorial framework, not a new benchmark or the result of our own experiment.

A demonstration can matter technically while leaving the last two questions open. Identifying this gap helps define the next investigation without dismissing progress or turning it into a sweeping prediction.

Autonomy must be evaluated across the whole system

Consider a hypothetical agent that receives a request, consults documents and changes an internal system. Its final answer may look correct even if a document was outdated or a tool received an unsuitable instruction. Evaluating only the final text misses important behavior.

A test should track the path: retrieved information, actions taken, dependencies and completion criteria. It should include cases where the best action is to ask for clarification or stop. Useful autonomy includes knowing when to continue and when to return a decision to a person.

A local evaluation can start small: select real tasks, define acceptable outcomes, log failures and repeat the test after relevant changes. Model selection then becomes a testable hypothesis within the architecture.

What this adds to the AGI discussion

Here AGI refers to the debate about systems able to learn and act with generality across domains. That working definition does not settle disagreements about the term or predict a date of arrival.

The scientific question is how capability transfers to new situations. The engineering question is what makes that capability reliable enough for a particular use. Keeping both explicit lets us recognize progress without confusing one signal with a conclusion about all intelligence.

This perspective connects to our work on agentic systems and evaluation and governance: building intelligence means understanding dependencies, observing behavior and making limits explicit.