Why AI Progress Now Happens Between Your Questions

The new unit of progress

AI & Robots introduces a useful distinction: a model generates responses, while an agent can choose steps, use tools and observe what happened. Its chapters on agents, trust and the future of work also emphasise that a job contains many tasks, and that connecting tasks is harder than performing one in isolation. This article develops those ideas into a question about the pace of change in October 2026. ^[1]

Imagine two assistants with equally convincing writing. One proposes a way to investigate a problem. The other, given suitable access, checks the relevant files, runs an analysis, finds an inconsistency and returns with a corrected draft. The difference is not the elegance of the last paragraph. It is the amount of the task that happened before that paragraph appeared.

That is an editorial example, not a benchmark. It suggests a more useful question to ask of a new system: how much well-defined work can it complete before you must intervene?

September brought infrastructure for longer work

On 10 September 2026, OpenAI introduced its Agents API in public beta. The announcement describes infrastructure for sessions in which agents work with files, run code, use tools and retain intermediate results. It also describes delegation to subagents. This is a developer product announcement, not proof that every deployed agent is dependable. ^[2]

Think of the difference between a talented visitor and someone equipped to work in an office. The visitor may have excellent ideas. An effective worker also needs the right documents, tools, workspace and a way to keep track of unfinished tasks. The analogy explains why improvements outside a model can change what a person experiences as AI progress.

Access is part of that change. An agent cannot check a private document simply because it can reason about documents. The system must connect the relevant source under appropriate permissions. Capability, access and approval are separate questions. ^{[1, 2]}

AI is helping with the work of building AI

A more consequential development happens inside research organisations. In a 6 September account, OpenAI described wider use of coding agents in its own research. It reported more experiments per active experimenter and increasing delegation of longer or more complex tasks. The company also noted that available computing resources had grown, so increased experimentation should not be attributed to agents alone. ^[3]

This creates the possibility of a feedback loop: tools help researchers implement and test ideas; some useful ideas improve later tools; those tools can then assist further work. This is our interpretation of the relationship, not a measured law guaranteeing accelerating intelligence.

OpenAI's own account says overall research progress need not match the growth of particular activity metrics. People still choose priorities and judge results, while other bottlenecks remain. More experiments are not automatically more discoveries. A failed idea tested quickly is still a failed idea, although learning that it fails can be useful. ^[3]

An agent-hour is not a human-hour

A system can run for a long time without producing a useful result. It can also finish a narrow task quickly while leaving the difficult decision untouched. To understand progress, separate activity from outcome.

For an illustrative supplier-comparison task, count more than the number of pages produced. Were the prices current? Did the comparison include delivery and exclusions? Can you trace each claim? How much correction did a person need to make? Would the result survive a second review? These questions turn an impressive-looking deliverable into an assessable one.

Anthropic's engineering guidance distinguishes fixed workflows from agents that decide how to proceed. It recommends evaluating complexity against the task rather than assuming a more elaborate system is always better. Its discussion also recognises costs and the possibility of compounding errors. ^[4]

That means progress may sometimes look surprisingly modest: one repetitive job completed consistently, with fewer handoffs and clear records. A dramatic demonstration is easier to notice. A reliable routine is easier to use.

What changes in ordinary work?

Consider a small team's weekly report. In a possible workflow, software collects permitted figures, highlights changes, drafts an explanation and prepares questions about unusual entries. People supply missing context, check the figures and decide what to communicate. This is a scenario for thinking, not a promise that every service can run it unattended.

The book's most useful workplace idea is that people, software and physical machines contribute different strengths. A better writing tool does not replace a delivery van. An agent that arranges a task does not perform its physical steps. Some jobs change through many small reallocations of work rather than one complete replacement. ^[1]

Start the discussion with a real bottleneck: what repeatedly consumes time, has an observable result and can be checked? Then ask which parts require judgement, a relationship or accountability. That framing is more concrete than asking whether AI will replace an entire profession.

Keep speed connected to control

When software moves from suggesting to acting, the consequences of a mistake can change. The book's trust chapter asks readers to verify important claims, limit access to what is needed and review consequential actions. NIST's AI Risk Management Framework likewise treats trustworthiness as something to address through design, development, use and evaluation. ^{[1, 5]}

For a new agent workflow, make the success condition explicit. Define what it may read, what it may change and where it should stop for review. A result should be easy to inspect, and a failure should be visible rather than buried beneath a confident summary.

The exciting part is the possibility of more useful work happening while you concentrate elsewhere. The hard part is making that work dependable. The next advance worth noticing may therefore be a cleaner chain of checked steps, rather than a louder claim about intelligence.