Are AI agent coworkers real? What the research says

Two pieces published within a day of each other this week arrive at the same conclusion from opposite directions: today's AI agent coworkers are a marketing label, not a workplace reality. MIT Technology Review argues that calling an agent a colleague quietly damages the human oversight that keeps the work safe. A survey paper from Tencent's Youtu Lab, covered by The Decoder, comes at it as an engineering problem and lays out what would actually have to change before an agent could be trusted to finish real work. One warns about the language. The other maps the missing machinery. Read together, they are a useful corrective to a year of "digital employee" hype.

The timing is not a coincidence. Since April 2026, Microsoft and OpenAI, along with Anthropic and Google, have all shipped tools for managing fleets of agents, and Nvidia's Jensen Huang has talked up "digital humans" joining the workforce. The vocabulary of hiring and reporting lines has spread faster than the underlying capability. Both reports push back on that gap, and the contrast between them is the interesting part.

The case against calling agents coworkers

MIT Technology Review's argument is not that agents are useless. It is that the employee framing changes how humans behave around them, and not for the better. The piece centers on research by Emma Wiles at Boston University, who tested how people review work depending on what they are told produced it. When the same output was attributed to an "agentic AI employee" rather than a plain chatbot, reviewers caught 18 percent fewer errors.

The mechanism is subtle. Framing the tool as a colleague seems to grant it a presumption of competence, the same social benefit of the doubt we extend to a human coworker. People relax their scrutiny. At the same time, Wiles found participants were 44 percent more likely to escalate questionable work to a manager when the agent was described as an employee, which eats into the time savings the agent was supposed to deliver. The framing both lowers the quality of oversight and adds friction, an unhappy combination.

This is already showing up in practice. The article reports that about 23 percent of 1,261 surveyed managers say they list AI agents on their organizational charts. Once a tool has a name like "Alex" and a box on the org chart, the line between software and staff blurs, and so does accountability. MIT Technology Review's sharpest point is that the employee metaphor "inverts our sense of who's in charge," making the human operator feel less responsible for output they are still on the hook for.

That accountability question stops being abstract in high-stakes settings. The piece names domains like healthcare and warfare, where a false sense of machine authority could do real damage. It points to one episode that was popularly blamed on Claude, where the outlet says the evidence actually points to a cascade of human errors rather than an autonomous machine decision. The lesson the author draws is that anthropomorphic framing makes it easier to assign blame to the tool and harder to fix the human process that failed.

Why the employee metaphor spread so fast

It is worth asking why the coworker language took hold at all. Part of the answer is commercial. A vendor selling a "digital employee" can anchor its price against a human salary, which makes a subscription that costs thousands of dollars a month look like a bargain next to a headcount. The metaphor is a pricing strategy as much as a description, and it flatters the buyer into imagining a tireless new hire rather than a piece of software with a usage bill.

The management tooling reinforces it. When the platform you use to deploy agents borrows the vocabulary of onboarding and performance, the mental model follows the interface. MIT Technology Review's concern is that this framing is not harmless decoration. The Wiles results suggest it measurably changes behavior, and the change runs in the wrong direction for anyone who cares about catching mistakes. A label chosen to sell a product ends up shaping how carefully that product gets checked.

There is a counterargument worth stating fairly. Naming a tool can make it easier to reason about and build workflows around, the same way naming a server or a script does. The objection is not to names as such, it is to a specific name that imports a whole set of social expectations the technology cannot honor. Calling an agent "Alex" and giving it an org-chart box is a different act from labeling a build pipeline, because it invites people to treat its output the way they would treat a trusted colleague's.

What honest framing looks like in practice is fairly mundane. An agent is a process a person runs and remains answerable for, the way an engineer owns the output of a script they wrote even if they did not type every line of it. Keep the human name on the work. Treat the agent's output as a draft that needs review rather than a finished deliverable from a trusted hire. The Wiles findings suggest this is not pedantry. The same reviewers who relaxed under the employee label would presumably hold the line if the work were simply presented as machine-generated and theirs to verify. The remedy the research points to is not better technology, it is refusing to let a marketing word quietly reassign responsibility.

What AI agent coworkers would actually require

The Tencent survey paper, written with collaborators at several Chinese universities, takes the opposite tack. Instead of asking whether we should call agents coworkers, it asks what it would take to make an agent reliable enough to deserve the comparison. Its answer reframes the whole problem. As the researchers put it, the central question is no longer how a model produces a better answer, but how it reliably turns intent into finished work.

That shift from answer quality to task closure is the heart of the paper. A chatbot succeeds when its reply is good. An agent only succeeds when the job is actually done, verifiably, in a real environment, with the side effects that implies. Those are different bars. The paper argues the industry has spent its energy on the first and underinvested in the second, which is why agents demo well and disappoint in production.

To structure the argument, the authors lay out a five-stage evolution. The chatbot era was one-pass text generation. The thinking-LLM era added extended inference compute, the approach behind models like OpenAI's o1 (2024) and DeepSeek-R1 (January 2025). The agent era bolted on tool calls and API access, but with fragile execution. The next stage, which the paper labels persistent-workspace agents, gives a system durable state to work in. The final stage it sketches is self-evolving AI ecosystems that improve their own tooling over time.

The labels matter less than the trajectory they describe. Each step adds something the previous one lacked: reasoning on top of fluency, then action on top of reasoning, then durable memory on top of action. The paper's point is that the leap from a clever chatbot to a dependable worker is not one more increment of model quality, it is a change in the kind of system being built. A bigger model at stage one is still stuck at stage one. That reframing is why the authors spend more time on environments and tooling than on raw model scores, and it is the part most directly at odds with the assumption, baked into much agent marketing, that the next model release will simply make agents reliable on its own.

From answers to task closure

The paper is blunt about why current agents are fragile, naming several structural bottlenecks. Agents perceive their environment in fragments rather than as a coherent whole. Their tool calls are stateless, so each action forgets the last. Execution breaks easily, and tasks often end half-finished because nothing enforces completion. These are not bugs to be patched one at a time, the authors argue, but symptoms of an architecture that was never designed for sustained work.

The benchmark history backs this up. The paper cites SWE-bench for software tasks, OSWorld for desktop operation, and WebArena for web navigation, where early systems built on GPT-4 (2023) completed only about 14 percent of tasks. Those numbers have climbed since, but the gap between a model that can describe how to do a job and a system that can reliably carry it out remains the central challenge. Task closure, not eloquence, is the metric that matters.

What task closure demands becomes clearer in the kind of work the benchmarks measure. SWE-bench asks a system to resolve a real software issue so that a project's tests pass afterward, which is a binary outcome no amount of fluent explanation can fake. OSWorld and WebArena ask an agent to operate a desktop or a website and reach a defined end state. In each case the grader does not read the agent's reasoning, it checks whether the world changed the way it was supposed to. That is a far harsher standard than a chatbot ever faces, and it is the standard the Tencent authors argue the field has to optimize for directly rather than hope falls out of better language modeling. The early 14 percent completion figure on web tasks is a reminder of how far the gap once ran, and how much of the recent progress has come from the surrounding scaffolding rather than the model in the middle.

This is where the two articles quietly agree. MIT Technology Review worries that we overestimate agents because the employee framing flatters them. The Tencent paper explains, mechanically, why that overestimation is dangerous: the systems genuinely cannot yet finish complex work without breaking. The marketing language and the engineering reality are pointing in opposite directions, and the safety risk lives in the gap between them. An agent that looks like a confident colleague but breaks on the third step of a five-step task is precisely the kind of tool people stop double-checking right before it fails.

Workspaces and reusable skills

The paper's proposed fix is concrete. Reliable agents need two things working together: a persistent workspace and a library of reusable skills. The workspace gives the agent durable files and sessions, with permissions and audit logs, so state survives across steps instead of evaporating after each tool call. It is the difference between an assistant with a desk and one that wakes up with amnesia every few minutes.

The skills half is equally important. Rather than re-deriving a procedure from scratch through ad-hoc prompting every time, an agent draws on modular, testable skill packages. The paper points to Anthropic's Agent Skills format, with its SKILL.md files, as an example of packaging a procedure into something reusable and inspectable. The promise is that workspace persistence plus reusable skills delivers performance gains that bigger language models alone do not, because the bottleneck was never raw intelligence, it was the lack of durable state and repeatable procedure.

The paper does not pretend this is free. It names a fresh set of failure modes that come with the architecture. A persistent workspace that holds credentials and tokens, plus sensitive files, widens the security surface every time the agent acts. Reusable skills can overfit to the cases they were written for or quietly go stale as the systems around them change. Keeping agents reliable therefore demands lifecycle management, workspace hygiene, and a way to roll back when something goes wrong. In other words, the machinery that makes an agent dependable is also machinery that has to be governed, which is not the picture a "digital employee" pitch tends to paint.

Where people still want a tool, not a colleague

Neither piece argues that automation is unwelcome. The MIT Technology Review article cites Stanford research finding that workers want automation, just selectively. In that work, law clerks happily accepted an agent that monitored case progress, while sales representatives rejected one that handled credit verification. The pattern is not anti-AI, it is contextual. People want help with the parts of a job they find tedious or low-stakes and want to keep control of the parts that carry risk or judgment.

That nuance is exactly what the coworker framing flattens. A real colleague is granted broad latitude across a role. A tool is given a specific, bounded job. The Stanford finding suggests workers intuitively prefer the tool model, because it lets them grant authority task by task rather than wholesale. Selling agents as employees asks people to extend trust they are not ready to give, and the Wiles result captures what happens when they do it anyway: oversight slips.

The engineering view and the labor view converge here too. The Tencent paper's emphasis on permissions and audit logs, with rollback, is the technical expression of the same instinct the Stanford workers showed. Bounded authority, with a record of what the agent did and a way to undo it, is how you get the benefit of automation without surrendering control. That is a tool, carefully scoped, not a coworker with the run of the place.

Naming the thing honestly

The shortest summary of both reports is that the word "coworker" is doing too much work. It promises reliability the systems have not earned and accountability they cannot bear. MIT Technology Review documents the human cost of the metaphor in measured drops in oversight quality. The Tencent paper documents the technical distance still to travel before an agent can close out real tasks without breaking. The two are describing the same gap from different sides.

That does not make agents a dead end. The Tencent roadmap is a credible account of how the technology improves, and persistent workspaces with reusable skills are already finding their way into real products. The honest position is that agents are powerful, scoped tools today and may become something closer to reliable collaborators later, once task closure is solved and the accountability questions have real answers. Until then, the most useful thing a company can do is drop the employee costume and treat its agents as what they are. The research released this week suggests the cost of pretending otherwise is paid in exactly the oversight you most need when an agent gets something wrong.