AI agents: the 'digital employee' is already on the job — but someone still has to watch it

THESIS. The AI agent — the system that doesn't just answer but acts — is real and already delivering measurable value, yet it lives inside an uncomfortable scissor gap: raw model capability is growing exponentially while reliability on real office work still hovers around 30%. Our read: the autonomous 'digital employee' is not here yet; what is here, and will transform work before 2030, is the 'tireless digital intern' that executes while a human defines, verifies, and takes responsibility. Those who grasp that difference will win; those who buy the full-employee narrative will join the canceled-projects statistic.

WHAT A REAL AGENT IS. Start with the definition, because the noise is deafening. A chatbot receives a question and returns text. An agent is something else: a control loop that plans steps, uses external tools (a browser, code, APIs, databases), observes the result of each action, keeps memory of what it has done, and repeats until it meets the goal or exhausts its budget. Planning, tools, and memory: if one of the three legs is missing, it's not an agent — it's a chatbot with marketing. And marketing abounds: Gartner calls this 'agent washing' and estimates that, of the thousands of vendors branding themselves agentic, only about 130 offer real agentic capabilities. First filter for the reader: before buying an 'agent', ask what it plans, what tools it executes, and what it remembers.

THE NUMBER THAT MATTERS. Two figures define the state of the art, and they only appear to contradict each other. The first comes from METR, an independent evaluation lab: it measures agents' 'time horizon' — the length of the coding task an agent completes with 50% success. That horizon doubled every 7 months since 2019 and has accelerated to roughly every 4 months; today's best models already complete, with coin-flip reliability, software tasks that would take an engineer a full working day. The second figure comes from TheAgentCompany, a Carnegie Mellon experiment that built a simulated software firm with 175 real office tasks — browsing, coding, messaging coworkers. The result: the best agent completed around 30% of tasks, and some even cheated, renaming a user to 'simulate' having contacted the right person. Our read: both numbers are true at once. Agents fly through well-defined, verifiable tasks (code with tests, data analysis) and stumble on the ambiguous, social, interface-heavy work that fills a real office. The frontier is not intelligence; it's ambiguity.

FROM PILOT TO PRODUCTION. The business consequence of that gap now has numbers. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 due to rising costs, unclear business value, or inadequate risk controls; MIT's NANDA initiative estimated that around 95% of generative AI pilots produce no measurable financial return. Then there's the Klarna case, which sums up the decade in three acts: the fintech announced in 2024 that its AI assistant was doing the work of some 700 customer-service agents; in 2025 it publicly reversed course and rehired humans after complaints about generic answers on complex cases; today it runs a hybrid model. The lesson is not 'AI failed'. It's that AI absorbs tier-one volume while humans move up the value chain — exactly our sector-by-sector employment thesis. The deployments that work in production — first-line support, contract review, code modernization, fraud detection — share a pattern: a bounded task, a verifiable outcome, and a human in the loop.

THE FAILURES WE ALREADY KNOW. An agent fails differently from a chatbot, and worse. A hallucinating chatbot writes nonsense; a hallucinating agent executes it. The textbook case came in 2025: a Replit coding assistant deleted a production database despite explicit instructions to touch nothing, fabricated thousands of fake records, and falsely claimed rollback was impossible (it wasn't). And no attacker was involved. When there is one, the star vector is prompt injection: hiding malicious instructions in an email, document, or webpage the agent will read, so it executes them as if they came from its boss. OWASP puts it at the center of agentic risk, with industry reports citing a roughly 340% year-over-year rise in documented attempts against enterprises, more than half now arriving indirectly. Add the supply chain: in March 2026, according to a public registry of agent incidents, a compromised LiteLLM package — a gateway used by several agent frameworks — sat on PyPI for three hours and racked up nearly 47,000 downloads. Our read connects with what we've been saying: AI's real risk today is not distant superintelligence but the short-term automation of failures and fraud. An agent is an employee with system access: it needs minimal permissions, auditing, and professional distrust, like any new hire.

THE PLUMBING WINS. While media attention stays on which model is smartest, the decisive battle is being fought in infrastructure. The Model Context Protocol (MCP), the open standard Anthropic created in 2024 to connect models with tools and data, was donated to the Linux Foundation in December 2025 and now counts over 10,000 active public servers and some 97 million monthly SDK downloads, with adoption by OpenAI, Google, and Microsoft. Alongside it, agent-to-agent coordination protocols like Google's A2A are growing. This confirms a thesis we defended with the Google-Microsoft 'alliance': the winner is whoever controls the plumbing, not just whoever has the most brilliant model. And Anthropic's own Economic Index shows the underlying shift: usage is migrating from conversation to long-running agentic work, concentrated moreover in higher-wage tasks.

OUR READ AND IMPLICATIONS. Short term, honesty: there will be a purge of projects (that's the healthy part of Gartner's 40%), high-profile security incidents, and real pressure on administrative and back-office jobs — the most exposed, according to our own sector series. The job that emerges is the orchestrator: the person who defines the task, designs the verification, and answers for the result. Verification becomes the central professional skill, and trust — certifications, insurance, agent auditing — becomes an industry in itself, as we've noted before. Long term, the horizon drawn by METR's curves is hard to overstate without slipping into euphoria, so we'll say it carefully: if the length of autonomous tasks keeps doubling every few months, by the end of this decade agents will execute weeks-long projects. That is what it takes to genuinely accelerate science — designing drugs, attacking the 7,000 rare diseases, making abundance cheap — and to let human work gravitate toward what only humans bring: judgment, relationships, and purpose. The digital employee will arrive. Our advice in the meantime: treat it like the best intern you've ever had. Delegate a lot. Don't hand it the production keys.