2 November 2026 to 8 November 2026
Faster is not the same as better
Claims that AI accelerates software delivery still rest largely on anecdotes. Reports of faster output are common; dependable figures for time saved, cost efficiency and quality trade-offs are not. That distinction matters because a coding agent can look impressively busy while moving expense and risk elsewhere in the process.
One coding task may now consume ten to fifty times more tokens as an agent plans, generates, tests and corrects its work. Token counts are not a business result, but neither are they irrelevant once the invoice arrives. The sensible unit of measurement is shipped software that works, alongside the full cost of producing and maintaining it. Until teams measure that combination, productivity claims remain optimism with a usage bill.
Agents inherit the web's old problem
The agentic web revives Tim Berners-Lee's 2001 vision of autonomous software navigating a machine-readable internet. The semantic web expected people and organisations to maintain rigorous structured metadata. They largely did not, so the current proposal hands the same unruly material to probabilistic models and hopes their tolerance for ambiguity counts as a solution.
That compromise becomes dangerous when agents move from finding information to acting on it. Agentic analytics already struggles with insight quality; weak reasoning can pass directly into decisions that people trust. Common Fabric's pitch for "human-centric agentic computing" offers a reassuring description, and co-founder Alex Komoroske brings more than a decade at Google, but neither phrase nor pedigree establishes cost, reliability or control. Human-centricity is meaningful only when a person can inspect, interrupt and correct the system before a plausible error becomes an operational one.
Control has a maintenance budget
Local models offer a clear answer to dependence on proprietary cloud systems. Ollama and LM Studio make on-device execution possible, with the promise of greater privacy and less cloud reliance. The exchange is equally clear: hardware, setup and ongoing technical effort. Autonomy is not free; it simply puts the work and the bill somewhere you can see them.
The same pattern appears in regulated deployment. The EMA has frameworks for assessing real-world health data across pharmacy records, electronic health records, imaging and laboratory sources, but those records still require substantial curation before training or regulatory submission. viAct's SOC 2 Type II attestation likewise confirms baseline controls for security, availability and confidentiality. That is useful admission paperwork, not proof that a video analytics platform will remain dependable under real load. Compliance can establish that an organisation has done the minimum properly; it cannot perform the deployment on its behalf.
Renaming risk does not govern it
President Trump's September executive order directs federal agencies to replace "artificial intelligence" with "Super Intelligence", without supplying a technical definition. Departments and contractors are already changing documentation. Whatever its political signalling value, the terminology tells buyers and operators nothing about capability, cost or failure. Nobody invoices by adjective, although procurement paperwork may now give it a sporting chance.
Australia's government is considering a different form of linguistic control: banning emotional intimacy in outputs from systems including ChatGPT, Gemini and Claude. A broad restriction could produce safer-sounding but emptier conversation without showing that harm has been reduced. Rowan Davies's warning about blandness identifies the central problem: tone is easier to suppress than risk is to define.
Across software delivery, autonomous agents, medical data and public rules, the test should be stubbornly concrete: what outcome improved, what did it cost, who can inspect it, and what happens when it fails? If a product, compliance badge or government instruction cannot answer those questions, its preferred terminology is decoration rather than evidence.
Keep the useful bits
Continue with AI Edge for the written briefing and glossary download.