The agents got the job

An AI named Mona ordered 120 eggs for a Stockholm cafe that doesn't have a stove. She also ordered 3,000 nitrile gloves for a shop averaging one customer per hour. Fast Company documented the "Hall of Shame" that Y Combinator-backed Andon Labs is proudly publishing as it hands real operational authority to a Gemini-powered agent managing ordering, scheduling, permits, and supplier contracts.

The funny startup story would be easy to dismiss. But zoom out from Mona's egg problem and the same structural bet appeared across four stories in 24 hours, at scales where the purchasing mistakes won't be charming.

Anthropic shipped 10 preconfigured agents to Wall Street: building pitchbooks, screening KYC files, reconciling ledgers, closing month-end books. Jamie Dimon stood on stage with Dario Amodei and called the infrastructure spend "worth the trillion-dollar investment." Sierra raised $950 million at a $15.8 billion valuation because more than 40% of the Fortune 50 already route customer interactions through its agents, from mortgage refinancing to insurance claims processing. And PayPal announced 4,500 job cuts over three years, with its new CEO framing AI automation as the company "becoming a technology company again."

The conventional reading groups these as adoption stories: another quarter, another wave of AI products. But what happened this week is categorically different from what happened last year. Last year, companies bought AI tools. This week, companies hired AI agents — gave them purchasing authority, client-facing roles, and operational responsibility over processes that were previously someone's job description.

The delegation gap

The distinction matters because tools and employees occupy different positions in an organisation's accountability structure. When a tool produces bad output, a human reviews it. When an agent has operational authority (signing supplier contracts, processing insurance claims, closing books) the question of who reviews what, and when, becomes genuinely unresolved.

Mona's egg order is a £40 mistake. An agent that miscategorises a KYC flag at JPMorgan is a compliance incident. An agent that incorrectly processes a mortgage refinancing at one of Sierra's Fortune 50 clients is a lawsuit. The capability gap between these agents is real, but the governance gap is identical: none of these organisations have mature frameworks for when an agent acts autonomously and gets it wrong.

This is the pattern the individual headlines miss. The funding, the job cuts, the product launches all point to the same moment: companies decided that agents are reliable enough to delegate to, and are working out the failure modes in production rather than in pilots.

There's historical precedent for this kind of institutional improvisation. When corporations first operated across telegraph networks in the 1870s, branch managers gained autonomous decision-making power faster than headquarters could write policies to constrain them. The legal doctrine of apparent authority, where a company is bound by what its agents appear authorised to do, took decades to catch up. We're entering a similar gap, except the agents aren't in branch offices. They're in the software.

For builders, the immediate implication is straightforward: the bottleneck isn't capability. Sierra hit $150 million ARR because the agents work well enough. The bottleneck is the scaffolding around the agent. The monitoring, the escalation logic, the audit trail, the answer to "who's accountable when this goes wrong at 2am?" If you're building agent infrastructure, that's where the value is migrating: not making agents smarter, but making delegation safe.

Mona will probably stop ordering eggs she can't cook. The harder question is what happens when the agents handling billions in financial transactions make their equivalent mistake, and whether anyone built the system to catch it before it compounds.

---