Work is a Spectrum

Gemini’s take on the spectrum of work

There’s a lot happening in the world of agentic work right now. Agent frameworks. Agent harnesses. Agent orchestration platforms. Agent protocols. Agent sandboxes. Agent registries. Agent skills. Agent on a stick.

Intertwined with the hype is an enthusiasm about AI agents that are now capable of doing certain kinds of work that previously required a human. That’s new and significant. But the conversation around it has gone sideways.

Hype from the ones selling the shovels

In March 2025, Anthropic CEO Dario Amodei predicted that AI would write “90% of code” within three to six months. That window has passed. He repeated the claim in early 2026, extending it to six to twelve months — and adding that AI would soon replace software engineers “end-to-end.” As of this writing, Anthropic has over 400 open engineering positions, some paying north of $400,000.

Sam Altman has oscillated between extremes. In March 2025, he said “maybe we do need less software engineers” and that mastering AI tools was “the new ‘learn to code.’” By May 2026, he was pushing back on AI job fears as “overly pessimistic,” saying OpenAI aims to “augment and elevate” human capabilities. The messaging shifts with the audience: alarm for investors, reassurance for regulators.

Meta’s chief AI scientist Yann LeCun responded directly: “Dario is wrong” — arguing that Amodei “knows absolutely nothing about the effects of technological revolutions on the labour market.”

The pattern isn’t new. Predictions in 2023 anticipated AI would replace up to 80% of developers by 2025. Instead, 95% of enterprise AI pilots have failed to deliver measurable returns, and software engineer job postings are up 11% year-over-year and rapidly rising. Starting salaries for CS graduates are up 7% year-over-year. Salaries across tech are projected to rise 8-10% in 2026. As CNN put it, the demise of software engineering jobs has been greatly exaggerated.

A decade ago, experts predicted robo-advisors would decimate financial adviser jobs. Those roles expanded. And just last week, both Anthropic and OpenAI launched enterprise AI services joint ventures — spending a combined $11.5 billion to embed engineers inside companies to help them adopt AI. Replacing engineers… by hiring more engineers.

Tasks, not jobs

The MIT Iceberg Index — the most rigorous analysis to date of AI’s workforce impact — helps explain why the predictions keep failing. Project Iceberg maps 32,000 skills across 151 million workers using O*NET occupational data to measure what AI can actually do. The finding: 11.7% of the US workforce (about $1.2 trillion in annual wages) is technically exposed to cognitive automation right now. But actual adoption? About 2.2%. The gap between “could be automated” and “is being automated” is enormous.

The reason is that LLMs automate tasks, not jobs. A job is a bundle of tasks, relationships, judgment calls, physical actions, and accountability. AI can draft a compliance report, but it can’t be held legally responsible for the report’s accuracy. It could analyze fermentation data against recipe targets, but it can’t taste the beer. It can generate code, but it can’t exercise actual judgment whether the feature should be built in the first place.

The Iceberg Index explicitly does not predict job losses or adoption timelines. It measures exposure — which tasks could theoretically be handled by AI. What happens with that exposure depends on firm adoption strategies, regulatory constraints, worker adaptation, and societal acceptance. These are human decisions, not technological inevitabilities.

Agents are new… work isn’t.

Humans have been doing work for as long as there have been humans. We’ve organized it, governed it, measured it, improved it. We’ve built guilds, factories, assembly lines, project management methodologies, quality standards, and regulatory frameworks — all in service of getting work done well.

With the advent of computers, we’ve had software helping us get work done for many decades now as well. Batch jobs, cron schedules, CI/CD pipelines, deployment scripts, API integrations, automated testing — deterministic automation that transforms data, calls APIs, and performs calculations reliably. Nobody calls a database trigger an “agent,” but it’s been doing work alongside humans for longer than most AI researchers have been alive.

AI is the newest worker on the scene of the Information Age, and it occupies the middle ground between the other two. Like a human, it’s non-deterministic — it needs instruction, direction, and its outputs require verification. Like code, it needs a runtime environment, context, and well-defined interfaces. It draws from both piles, which suits it well for a different set of tasks than the others.

Three worker types. Not one replacing the others — all three participating in a spectrum of work.

The spectrum is a choice

Even where AI can handle a task, that doesn’t mean it should — or that people will want it to.

The Stanford 2026 AI Index Report found that only 10% of Americans are more excited than concerned about increased AI use in daily life. 64% think AI will lead to fewer jobs. Gen Z — the heaviest AI users — are growing less hopeful and more angry about it as they feel AI is something being inflicted upon them, and not for their benefit.

Industries and professions will make different choices about where AI participates.

Ethical concerns. Creative professionals are pushing back against AI-generated content. Apple TV’s “Pluribus” carries credits stating “this show was made by humans.” iHeartMedia launched a “guaranteed human” campaign. Human-made badges increase perceived value by up to 3x in blind tests.

Environmental concerns. Massive energy demand, high water consumption for cooling data centers, and electronic waste from specialized hardware that’s obsolete within 18-24 months — today’s AI agents are very costly for the environment. Those that care about sustainability may limit or avoid AI usage altogether.

Legal liability. AI can draft a contract. It can’t sign one. When the SEC audits your financial statements, a human is answering the questions. When a patient sues over a misdiagnosis, a human is in the deposition. Accountability is a human function, and the legal system isn’t changing that anytime soon.

Copyright. Deezer reports that 44% of daily music uploads are now AI-generated — nearly 75,000 tracks per day — and has demonetized 85% of AI-generated streams. Bandcamp prohibits music generated “wholly or in substantial part” by AI. The US Supreme Court declined to extend copyright protection to AI-generated works in March 2026. The New York Times, Daily News, and Chicago Tribune have all sued AI companies over unauthorized use of their content. Industries where IP provenance matters — publishing, music, pharmaceutical research — are already restricting AI involvement to maintain clear ownership chains.

Preference and purpose. Work isn’t just economic output. It’s how people find meaning. Research consistently shows that employed people report higher life satisfaction than unemployed ones, even controlling for income, because work provides structure, identity, connection, and purpose. Humans are happiest when meaningfully engaged and even challenged — lasting satisfaction comes from mastering skills, contributing to community, and the sense of being needed.

As Ky Decker put it in his post on AI burnout: “Things that are worth doing are worth doing well… Ideas are common; effort is not… There are no shortcuts.” The point of a code review “is not simply for good code to make it into a codebase, but to build institutional knowledge as people debate and iterate.” Automating the output misses the point when the process itself is where the value lives.

The assumption that everyone wants AI to do their work for them is a Silicon Valley/Big Tech projection, not a maxim. Many people find fulfillment in the doing.

What we actually need

So we have three worker types, a spectrum of work that spans all of them, and a world where different industries, organizations, and individuals will make fundamentally different choices about where each type participates. The question isn’t whether AI will do work — it already does. But agent-only frameworks only address one worker type; they aren’t addressing the greater issue of how to get all three types working well together.

What we need is work infrastructure that treats the full spectrum as a first-class concern, from fully manual to fully automated and everything in between.

Three worker types as peers. Humans, AI agents, and code workers participating in the same governed processes. Not AI bolted onto human workflows. Not humans just observing from the outside. All three as first-class participants.

Governance that’s structural, not aspirational. Quality gates, independent review, audit trails, containment — enforced by architecture, not by hoping the agent follows its system prompt. When work matters, “trust the AI” isn’t a governance model.

Knowledge that compounds. The 100th run of a process should benefit from everything the first 99 taught the system. Not because the AI got smarter, but because every worker — human, agent, and code — contributed observations, decisions, and outcomes to a shared understanding.

The right worker for the right task. Each worker type has their respective strengths. With an understanding of the task at hand, the infrastructure should facilitate matching the work to the worker type best suited for it.

Respect for the choice. The infrastructure should work just as well with no AI at all, or any mix in between. The business owner, the operator, the practitioner — they decide where AI participates. The platform supports whatever they choose.

Work is a spectrum

The agentic work conversation has been dominated by this narrative that AI agents are the future, everything should be rebuilt for agents, and humans are the bottleneck to be eliminated.

That narrative is wrong — not because AI isn’t powerful, but because it mistakes one worker type for the whole workforce. AI agents are a significant addition to the team, not a replacement for it.

Work is a spectrum. The infrastructure we build going forward should reflect that.