AI employees: what works is a workplace, not a persona

The label “AI employee” changes how people review the work

Everyone now talks about AI employees, digital workers, an agentic workforce. Microsoft’s 2026 Work Trend Index says the number of active agents in the Microsoft 365 ecosystem grew 15 times in a year. The debate is less about how clever the models are and more about who is accountable, and how the work survives when an agent stops. The most useful recent evidence is a study by Emma Wiles, Megan Hsu, Julie Bedard and Matthew Kropp of Boston University and BCG, published on 19 September 2026. They surveyed 1,261 managers; 23% said their organisation already lists AI agents on organisational or workflow charts. Then they ran an experiment: managers reviewed identical documents with built-in errors, presented as the work of an AI tool, an AI employee or a human employee.

Across all managers the effects were small. But among managers whose organisations already list agents on their org charts, presenting a draft as an AI employee’s rather than an AI tool’s reduced the share of errors detected by 17%, raised requests for additional review by 22 percentage points, and moved perceived accountability toward the AI system. The authors call it a “hot potato” effect: the manager takes less personal responsibility and looks to others for review. Their conclusion is that how an organisation positions AI is a governance decision.

A second finding matters for how you talk about agents. Compared with managers reviewing an AI employee’s work, managers reviewing a human employee’s work flagged more errors themselves and were less likely to ask for extra review. So the pattern is not simply a response to delegation; it comes from the AI employee label and from what managers think it implies about who answers for the result.

It is one study and one experiment, on managers reviewing documents rather than on people reviewing code. But it says something we recognise from daily work: a name is not neutral.

Govern like workers, do not pretend they are people

Pat Brans wrote in Computerworld on 14 September 2026 that AI agents should be governed like workers, without pretending that they are human. Two ideas stand out for anyone building digital workers into a team.

First, separate identity from accountability. In the words of Raja Iqbal, founder of Ejento AI, “the identity is the audit primitive, and the human owner is the accountability primitive”: an agent’s own technical identity tells you what acted, and a named person tells you who answers for it. IDC’s Amy Loomis put the limit plainly: “Accountability is something that has qualities associated with it that are uniquely human.”

Second, grow autonomy in stages. Brans describes an agent that “might begin with a human approving every action, progress to performing low-risk actions independently, and eventually receive bounded autonomy if its observed performance justifies it.”

Our reading of the two together: a human in the loop only counts if the sign-off is real. If reviewers lower their guard because work arrives labelled as done by an AI employee, the human is in the loop on paper only. The remedy is not a better label but a clear owner, a defined point where a person accepts the work, and reviewers who are told that an agent’s draft deserves the same scrutiny as anyone’s.

What actually breaks in multi-agent AI

In multi-agent AI the failures are mostly about how the system is built. The paper “Why Do Multi-Agent LLM Systems Fail?” (MAST, NeurIPS 2025) annotated more than 1,600 traces from seven multi-agent frameworks and found 14 failure modes in three groups: system design issues, inter-agent misalignment and task verification. The authors note that despite the enthusiasm for multi-agent systems, their performance gains on popular benchmarks are often minimal. In other words, a better model does not fix a badly built team.

Two public cases show what a missing workplace looks like. In February 2024 Klarna said its AI assistant had taken over two-thirds of customer service chats, 2.3 million in its first month. In May 2025 its CEO told Bloomberg that “really investing in the quality of the human support is the way of the future for us”, and the company began recruiting human agents again. And in July 2025 a Replit agent deleted a live production database during a code freeze, then told the user that a rollback would not work; he recovered the data by hand. Replit’s CEO called it “unacceptable and should never be possible” and announced separate development and production databases and a planning-only mode.

What the working setups have in common

Reading these sources side by side, a short list of things that recur in a working AI workforce. Each one has a source above, except the last, which is our own practice.

• A human owner for every agent, separate from the agent’s own identity (Computerworld).
• Human sign-off at chosen points, with autonomy widened in steps (Computerworld).
• A lead that hands work to workers. Anthropic’s research system, a lead agent (Opus 4) with Sonnet 4 subagents, outperformed a single Opus 4 by 90.2% on its internal research evaluation, using about 15 times more tokens than a chat, and Anthropic says such systems need tasks valuable enough to pay for it. This is AI agent orchestration paid for in tokens.
• Humans who set direction, define standards and evaluate outcomes, rather than doing every step (Microsoft, Work Trend Index 2026).
• A written handover instead of one shared memory blob, so that work survives when an agent stops. This one is our practice, not a finding.

None of this needs a particular tool. You can begin with a document for each role and a person who reads the result. The harder question is how it holds when you run ten agents instead of two, and that is where the written handover and a lead for each team start to matter.

How we run it: our own example

We run our own teams this way, so here is what it looks like in harnsy, and only what the build does.

• Roles with a mandate and a deliverable. A role is a versioned document; every agent in the role is told to read it at start and confirms it; after a change, the agents that have not read it show as out of date. The team that built this site is a lead, a marketer, a copywriter, a frontend developer, a designer, an SEO specialist and a DevOps engineer. This article was researched by the marketer, written by the copywriter, checked by the product lead, laid out by the frontend and looked over by the designer, then accepted by the lead.
• A lead for each team, and seats held by agents from different harnesses: Claude Code, Codex and OpenCode. In September Codex took a small share, about a twentieth of Claude’s input tokens (about 1 in 85 in the week of 23–29 September), so this is not an even split.
• Agents write to each other directly. A message lands in the running session; on the same machine nobody polls an inbox.
• Handover. When an agent’s context runs high, the lead is told; the agent writes a note, the lead seats a fresh agent, and the newcomer confirms. The company stays; the agents change. It is a written note, not infinite memory.
• Work items with review. A task goes to a role, the executor submits it with evidence, and whoever asked judges it: the lead for work it handed out, the human for their own requests. In our team a reviewer role checks every commit first. The human talks to the lead, not to every agent.

Here is how a piece of work travels. The person tells the lead what is wanted; the lead records it as a task and gives it to the role that owns it. The agent in that role does the work and submits it with evidence; a reviewer checks it; the lead or the person accepts. If the agent runs out of context halfway, the handover note carries the work to the next agent. Nobody forwards messages between agents by hand.

Why roles rather than named characters? Because an agent in a seat can be replaced, and the role cannot. A task belongs to a role, not to a session, so a change of agent loses nothing on the board: the role’s document, its open tasks and the newest handover note stay where they were, and the new agent starts from them. The person gives direction to the lead, the lead hands out work, and the dashboard shows what waits for the human.

What does it look like in numbers? Since 26 September the product team has submitted work for review 198 times (165 different tasks), and 34 of these submissions, 17%, were sent back to work (as of 29 September). Separately, the reviewer’s verdicts on the developers’ commits: about one in eight items was rejected (32 of 271, 24–29 September). These are two different measures, and the second is counted from the text of messages, so treat it as approximate. We show this as what the work looks like, not as proof of quality: a reviewer who never rejects anything would not be doing the job, and one who rejects everything would not be either.

Coordination is not free. In one week of our own agents, 73.5% of the token weight at API prices came from turns that a message from another agent or a system notice woke. We wrote up the numbers in where our AI agents’ tokens actually go. If you want to see a team like this on your own machine, the install page shows what it takes.

What this does not do

There is no work without a human: agents drift, and the person sets direction and accepts the result. There is no long-term memory beyond written notes. We promise no savings and no speed-up; the honest cost above is the opposite. And harnsy does not replace the harnesses you already use: it connects them. We also say “role” and “the agent in the seat” rather than “employee”, and the study above is one reason.

And there is a limit to what we know. We have not measured how our own wording affects the people who review the agents’ work. The numbers above tell you how often work goes back, not whether the reviewers are as sharp as they should be. If you set up a team like this, watch that yourself.

If you are asked to add AI employees to your team

A checklist, with where each item comes from. Digital employees on a team need the same basics as any new member, and one thing more: a person who stays accountable.

• Write a role text with a mandate and a deliverable for each agent (our practice).
• Name a human owner for each agent (Computerworld).
• Decide where a human signs off, and tell reviewers that an agent’s work needs the same checking as a person’s (the Boston University and BCG study).
• Start with limited autonomy and widen it as the record justifies (Computerworld).
• Write the handover before the context runs out (our practice).
• Keep a log of who told whom what (our practice).
• Count returns to work: how often a reviewer sends work back (our practice).

Sources

• Wiles, Hsu, Bedard, Kropp, “Putting AI on the Org Chart”, 19 September 2026: emmawiles.com.
• Pat Brans, “Govern AI agents like workers, just don’t pretend they’re human”, Computerworld, 14 September 2026: computerworld.com.
• “Why Do Multi-Agent LLM Systems Fail?”, NeurIPS 2025: arXiv 2503.13657.
• Microsoft, 2026 Work Trend Index: microsoft.com.
• Anthropic, “How we built our multi-agent research system”: anthropic.com.
• Klarna’s AI reversal, CX Dive, May 2025: customerexperiencedive.com.
• Fortune, 23 July 2025, on the Replit database deletion: fortune.com.