Agent boss

An agent boss is Microsoft's term for a person who builds, delegates to and manages AI agents. It comes from Microsoft's Work Trend Index of April 2025, which predicted that "every employee becomes an agent boss" as organisations move to teams of people and agents. The term is useful because it names a real change: more people will direct software that takes actions, not just software that answers questions. It is also misleading, because the title arrives without the thing that makes management work, the ability to judge the work being done. Microsoft's own 2026 report did not use the phrase, and when it asked AI users which human skills matter more as AI takes on more work, the top answer, from half of them, was quality control of AI output.

Where the term comes from#

Microsoft introduced it in 2025: The Year the Frontier Firm Is Born, published on 23 April 2025 and based on a survey of 31,000 knowledge workers in 31 markets. The report sets out three phases. In the first, AI acts as an assistant. In the second, agents join teams as "digital colleagues" taking on tasks at human direction. In the third, "humans set direction for agents that run entire business processes and workflows, checking in as needed", an organisation Microsoft describes as AI-operated but human-led. Firms may be in all three phases at once.

Around the agent boss the report built a vocabulary: the Frontier Firm, the "human-agent ratio" as a new metric, and a "work chart" of teams formed around goals in place of the org chart. A WorkLab article in May 2025, How to Be an Agent Boss, broke the role into three: build an agent for a specific problem, delegate with the right amount of human oversight, and manage it by giving feedback and reviewing its performance.

What the survey found#

The figures show a gap between leaders and everyone else. 67 per cent of leaders said they were familiar with agents, against 40 per cent of employees. 36 per cent of leaders expected to manage agents in their own jobs, against 21 per cent of employees. Asked what their teams would be doing with agents within five years, leaders said building multi-agent systems (42 per cent), training agents (41 per cent), redesigning processes with AI (38 per cent) and managing agents (36 per cent). 81 per cent expected agents to be moderately or extensively part of their AI strategy within 12 to 18 months.

These are expectations, gathered in early 2025 by a company that sells agents. They describe what leaders believed was coming, not what had happened.

What changed by 2026#

Microsoft's 2026 Work Trend Index, published on 5 May 2026 from a survey of 20,000 people who use AI at work, kept the Frontier Firm and did not use "agent boss", the human-agent ratio or the work chart. It reported that 86 per cent treat AI output as a starting point. Asked which human skills become more important as AI takes on more work, 50 per cent named quality control of AI output, the top answer, ahead of critical thinking at 46 per cent. The emphasis had moved from directing agents to checking them.

Others had moved the same way. Bain's operating model paper of May 2026 used the term in quotation marks: "employees become 'agent bosses' overseeing digital labor alongside their own work". A Harvard Business Review article in February 2026 argued for "agent managers" as a distinct, specialist role. Brien Posey, writing in Redmondmag in December 2025, described the agent boss as "an already overworked employee who is now tasked with babysitting bots".

The sharpest finding came from an experiment reported in Harvard Business Review in May 2026, Why You Shouldn't Treat AI Agents Like Employees: treating agents as if they were people "reduced individual accountability, increased unnecessary escalation, lowered review quality". The management metaphor itself weakened the oversight.

Can people supervise agents well?#

What agents can do is growing fast. METR measured the length of software tasks that frontier models can complete and found it "doubling approximately every seven months since 2019" (Kwa and colleagues, March 2025); see the METR time horizon. But reliability on ordinary office work is still low. In Carnegie Mellon's TheAgentCompany benchmark, the best agent completed 30 per cent of simulated workplace tasks on its own. Salesforce's CRMArena-Pro found around 58 per cent success on single-step business tasks, falling to about 35 per cent over several turns, with "near-zero inherent confidentiality awareness". Gartner predicted in June 2025 that over 40 per cent of agentic AI projects would be cancelled by the end of 2027.

So the agent boss is asked to supervise a worker that is improving quickly, fails often, fails fluently and does not announce when it is out of its depth. That is a harder supervisory job than managing a person.

Where this sits in my own argument#

Microsoft files reviewing an agent's performance under managing it. I would put it first, because judging the work is the whole job. A manager who cannot tell good work from bad is forwarding, whatever the title says. Most good managers learned to evaluate by doing the work first. The agent boss model asks people to manage work they may never have done, and the organisation stops creating the experience that would let the next generation do it. That is the problem of supervising work you cannot do, and the reason I treat oversight readiness as a capability to build rather than a title to hand out.

The agent boss also blurs who answers. If an agent sends the wrong email to a customer, was it the agent boss, the person who built the agent, or the vendor? No court or statute yet treats a model as answerable; liability sits with people and companies. In custody terms, every agent action needs a named person who holds it. In drift versus design terms, the human-agent ratio is the wrong metric to design around. The right ones are which decisions an agent may take, who can stop it, what the people supervising it must remain able to do themselves, and how anyone would know if it went wrong.

What an organisation can do#

Give every agent a named owner. One person who can say what it is allowed to do and will answer when it goes wrong. See who manages AI agents.

Write the limits before the agent runs. Which systems it can reach, which actions need approval, and when it must stop. See what your agents can reach.

Train judgement of output. Prompting is the easy part; the scarce skill is judging what comes back. Build it by having people do the work by hand often enough to know what good looks like.

Keep agents off the headcount. A human-agent ratio rewards deploying more agents. Measure outcomes, error rates and the time people spend checking instead.

September and October 2026: the consultancies measure the gap#

Five of the large consultancies published on agents within seven weeks, and read together they describe the agent boss arriving faster than the means to supervise one. BCG's Applied AI Index (30 September), a survey of more than 1,300 senior executives, found that 42 per cent of companies expect their agents to act autonomously by 2030, "deciding without human approval", while "only 5% of companies have the full set of critical controls in place today". Companies with all six of the controls BCG names generated three times as much value from agents.

A week earlier BCG's The Authorization Gap (22 September) reported that 35 per cent of organisations already have agents in production and 44 per cent are planning deployments, and put the oversight problem in one sentence: "A reviewer who receives a vague request after the agent has done the hard work is a rubber stamp." PwC's governance paper (11 September) went further, saying that human-in-the-loop supervision "is no longer sustainable where agents increase in number, pace, and complexity", that "every agent needs a business owner who is accountable for the agent's decisions and actions", and that the board should oversee whether the organisation has effective governance for agentic AI.

Deloitte's AI agents in the human workplace (10 September) reported that nearly 70 per cent of respondents to a Deloitte survey named "an inability to trust and govern agents" as a major barrier, and quoted a financial services executive: "The machine will never be accountable." And Bain's survey of 155 software companies for its Technology Report (29 September) put a number on the new job: time spent directing AI agents rose from 4 per cent two years ago to 23 per cent now, and is expected to reach 34 per cent in two years.

Read together, the five reports point to the same unmet need: a named person who owns each agent and can judge and stop what it does. That is the agent boss with the judgement put back in.

What this does not show#

The agent boss figures are survey expectations from 2025, collected by a company selling the technology, and Microsoft's later report moved away from the term. The benchmarks measure agents on simulated tasks, which may understate or overstate how they behave inside a real organisation with real tools. The Harvard Business Review experiment on treating agents like employees is reported here from its published summary. None of this shows that people cannot manage agents. It shows that the skill the role depends on, judging work, is the one most at risk if people stop doing the work themselves.