The hard question in AI right now is not capability — it is who has agreed to let the agents in, and on whose terms.
🔍 THE BOTTOM LINE
The agent era arrived faster than the agreement: DeepSeek shipped an agent its own coding agents largely built, Greenhouse surveyed a hiring market nobody on either side thinks works, Canada attached a signed job count to an equity cheque the same week OpenAI’s investors got selective — and the direction is one-way: capability is not the bottleneck any more, mutual consent is.
Agents shipped; nobody signed
On Saturday, DeepSeek published dsh-superagent — an autonomous coding agent that watches a repo and opens a pull request when asked — MIT-licensed, 6,930 stars and 693 forks this morning. One README line does more work than the feature list: “This code grew out of our team’s day-to-day development work, with much of the engineering done by coding agents.” Agents writing the plumbing that manages agents is the recursive workflow labs have sold for two years. It shipped over a weekend.
Two days earlier, on 1 November, TechRadar reported a Greenhouse survey of 1,100 UK and Irish job seekers: 81% rejected without ever speaking to a human, 42% sat an AI-run interview, and younger candidates now rewrite their CVs by hand to avoid “sounding like AI” — because everyone uses the same tools, at both ends of the table.
And five days before DeepSeek’s drop, on 29 October, an American detection vendor launched a model that classifies website visitors by how they move their mouse — pointer paths, scroll rhythm, typing cadence — to decide, client-side, whether the visitor is a person, a bot, or an agent dressed as one.
Four moves in five days. Capability shipped; consent didn’t. Not one came with the other party’s agreement — the workforce integrating the output, the candidate behind the screener, the person being impersonated.
The missing contract, being written clause by clause
If the agent era has a contract, moves like these are its drafting sessions.
Agents must announce themselves. A detection vendor now sells client-side agent detection trained on cursor and keystroke behaviour — self-identification is moving from etiquette to market segment, protocol drafts already in progress.
Agents must know whom they serve. DeepSeek scoped its agent with rules about who can ask, what it may do, when it may report. “An agent that can act for a whole team can also waste a whole team’s attention” — the design principle is consent mechanics.
Humans keep a veto. Amazon’s guideline, published 27 October, says that when AI doesn’t help your work, that is “a finding, not a failing” — while The Register’s read was that “not adopting AI or increasing output metrics” reads as falling behind. The consent test is not a memo; it is what an employer does with the engineer who documents AI underperformance. Proton’s cancelled re-interviews had already shown the veto is real: adoption imposed through fear can be reversed by organised workers.
The money is repricing to match
On the capital side, OpenAI cut its fundraise target from about $30 billion to $20 billion, $11 billion banked at the time of reporting; The Wall Street Journal’s word for the response from some large investors was “tepid.” Investors are repricing how much AI infrastructure is worth before it can show a return.
Canada moved the opposite way: on 29 October, the government’s release announced up to C$1.6 billion of equity in Cohere — against commitments of at least 1,000 new high-paying Canadian technical jobs by 2030 and C$1.7 billion of Canadian computing infrastructure. Equity makes the deal a revenue-sharing position, not an unconditional subsidy. New Zealand has not produced a named-promise instrument like it.
Where this lands in New Zealand
New Zealand sits on both sides of the gap: our firms are exactly the size the scoped agent model fits, while our graduates sit inside the same screening cycle as the UK cohort — 1,000 AI-written applications against one NZ job ad, recruiters saying the scarce edge is being identifiably human.
Three practical notes, all within existing policy settings.
For NZ businesses: the scoped-agent design is the safest deployment shape — MIT-licensed, runs against your own repos, needs no datacentre — but a pilot needs named authorisers and an audit trail, or the first leaky agent will set adoption back a year.
For job seekers: the Amazon guideline says it plainly — document where AI does not help your work. Showing where AI underperforms demonstrates the judgement the market pays for: the documented-judgement premium. Hand-editing a CV is the same instinct: being identifiably human, on purpose.
For government: Canada has written the first fully-specified answer in a peer small-country sovereign-AI debate — named job count, named infrastructure, stage-gated, public reporting. If NZ adopts any version of that instrument, it should be at least as verifiable.
❓ FAQ
-
Isn’t “agents shipped faster than consent” just how technology always arrives? Partly — but the last comparable general-purpose shift, the web, grew inside a visible rulebook: robots.txt, search ranking, eventually regulation. Agent infrastructure is deploying with none of those; the negotiation happens after capability ships.
-
Where does DeepSeek’s agent fit in the story? It is the proof of pace: a frontier lab shipped an agent mostly built by agents, over a weekend, MIT-licensed, on GitHub inside 24 hours, at 6,930 stars within a day. When capability moves that fast, the binding constraint shifts from the technology to the consent structure around it.
-
Is the Gen Z CV story really part of the same pattern? Yes — the screeners’ side of the same loop: platforms filter for identical AI-polished signals, so candidates hand-edit their CVs to avoid machine-pattern phrasing. Same consent problem, one applicant wide.
-
What can a small NZ company actually do? Two cheap things: run agents only with named authorisers and an audit trail, and state plainly which hiring stages humans review. Both beat the mutual screening mess Greenhouse measured.
-
Where does this go in 12–24 months? Tools for proving humanity — in the browser, the hiring funnel, code review — become product categories, not benchmarks (agent detection is already there). And DeepSeek-style scoping contracts become normal workflow: the org chart gains a row for “who authorises agents,” the way it gained one for API keys.
🔍 THE BOTTOM LINE
For two years the transition was measured by what models can do; that measure ran out this week. Everything shipping now ships with capability already in the box, and the open question is whether the other party agreed. Capability without consent builds fragile; capability with consent builds to last.
