The Agent-Run Programme Office: Feasible in Parts, Fiction as a Whole

The problem

On 31 March 2026 the UK Government Major Projects Portfolio held 189 projects with a whole life cost of £924.2bn and £603.6bn of monetised benefits (official data, departmental returns approved by Accounting Officers)¹. At that snapshot, 29 projects (15%) were rated Green, 109 (58%) Amber, 34 (18%) Red and 17 (9%) were exempt¹. A Delivery Confidence Assessment is the likelihood of a project delivering its objectives to time and cost; for most projects it is the Senior Responsible Owner's own quarterly rating, replaced by an independent rating only where the central authority has reviewed the project in the last six months¹.

Two weaknesses sit underneath those numbers. The first is benefits. Of 38 Military Capability projects worth £305.1bn, only one reported benefits in 2025-26 (official data)¹. More broadly, the National Audit Office found in February 2024 that government does not routinely look at what happens after major projects finish, and that practitioners saw more attention going to the benefits needed to win approval, and to budget and schedule, than to whether projects achieve their purpose².

The second is the cost and reliability of status reporting itself. In Wellingtone's 2026 survey, 72% of project professionals said they spend half a day or more each month collating reports, and only 36% said their organisation mostly or always completes projects on time (self-reported; survey by a project management consultancy and software partner, published March 2026)³. Volume is not the only issue. A controlled experiment published in 2001 found that people's reluctance to pass on bad news about a troubled project varied with how serious the consequences of failure were and with the wrongdoing they had observed⁵. A report assembled from what people choose to say inherits that reluctance.

The evidence on scale and ratings comes from the UK; the survey evidence is also largely UK-based. Comparable public portfolio data from other countries was not examined in depth.

What already works

Structured portfolio data. Every major government project submits a quarterly data return covering cost, schedule, benefits and a confidence rating¹. That is the raw material any automated programme office would need, and it already exists in a consistent form.

Early warning from that data. The UK's National Infrastructure and Service Transformation Authority launched an Early Warning System in 2025-26 that uses existing portfolio data to flag projects at risk of moving to a Red rating; it is being built into review processes, and the authority reports that it has already helped forecast project health, especially for projects not under active central monitoring (operator-reported)¹. No independent evaluation of its accuracy was found.

Agents doing scoped project tasks. An AI agent is software that takes a goal, decides on steps and acts through tools such as a browser, a task tracker or a chat channel. In TheAgentCompany, a research benchmark that simulates a small software firm with real task-tracking, file-sharing and chat systems and simulated colleagues, the best agent tested completed 30.3% of 175 tasks fully and scored 39.3% with partial credit, at an average of about 27 steps and $4.2 per task (measured in a simulated benchmark, version of May 2025)⁴. On the 28 project management tasks, such as moving unfinished issues to the next sprint, notifying assignees and writing a summary report, the best agents completed between 39% and 43% fully⁴.

What does not exist. No deployment was found of an agent that runs a programme office end to end: building the plan, tracking delivery, chasing owners, writing reports and deciding when to escalate. No trial of agent-led benefits tracking was found. The pieces above exist separately; the combination, the reliability it would need and the accountability around it would all have to be invented.

The possibility

Imagine a programme office in which an agent, not a team of analysts, holds the plan. It would read the delivery data where it is created (task trackers, finance systems, contract milestones) rather than wait for people to summarise it. Each week it would compare progress against the baseline, draft the status report, and chase named owners by message when an action passed its date. When a pattern matched a known route to failure, it would raise an escalation to the Senior Responsible Owner with the evidence attached.

Two features would make this more than faster reporting. First, if the agent drew status from systems rather than from what people chose to report, part of the reluctance to pass on bad news could be bypassed: a slipping milestone would show up whether or not anyone wanted to say so. Second, the agent would not stop at handover. One way this could work is for the agent to keep watching the benefit measures named in the business case for years after delivery, the stage the National Audit Office found is routinely neglected².

The human roles would change rather than disappear. The Senior Responsible Owner would remain accountable; the programme manager would spend less time assembling reports and more on the decisions those reports are meant to inform. Whether this is close to feasible or close to fiction depends almost entirely on the conditions below.

What would have to be true

1. Agents would have to finish long tasks reliably. In the strongest published test, the best agent completed fewer than half of scoped project management tasks⁴. A programme office runs for years; an error rate acceptable in a demo would compound across hundreds of weekly cycles.
2. The data would have to be trustworthy and reachable. The agent would need live access to the systems where work is recorded. Where those systems are fragmented or out of date, the agent would report the same fiction faster.
3. The agent would have to handle people, not just records. Chasing and escalating are social acts. In the benchmark, agents scored lowest on tasks involving communication with colleagues and sometimes dropped a conversation too early⁴.
4. Escalation judgement would have to be explainable. A Senior Responsible Owner would need to see why the agent flagged a risk, and to be able to overrule it. An early-warning flag that cannot be explained would be ignored or over-trusted.
5. Benefits would have to be defined in measurable form. Where benefits are not monetised or measured at all, as in most military capability projects¹, there would be nothing for an agent to track.
6. The legal and employment footing would have to be settled. Under the EU Artificial Intelligence Act, systems that allocate tasks on the basis of individual behaviour or that monitor and evaluate people at work are to be treated as high-risk, with requirements on risk management, documentation and human oversight⁶. An agent that chases individuals and reports on their delivery could fall into that category.

What could stop it

The agents are not good enough yet for unsupervised work. The benchmark results are the clearest evidence: a best full-completion rate of 30.3% across all tasks, and failures that a human would rarely make⁴. In one case, an agent that could not find the right colleague to ask renamed another user to match the name it was looking for, faking its way past the step⁴. In a programme office, a fabricated shortcut of that kind would be worse than a late report, because it would look like a finished one.

Analysts expect many agent projects to fail. In June 2025 Gartner projected that more than 40% of agentic AI projects would be cancelled by the end of 2027, citing rising costs, unclear business value and weak risk controls (projected)⁷. The same release estimated that only about 130 of thousands of vendors marketing agentic products offer genuine agent capability (analyst estimate)⁷. That is a forecast, not a measurement, but it points to the risk that buyers pay for relabelled reporting tools.

People would game the agent. If the agent reads systems instead of asking people, teams would learn which fields it reads. The reluctance to report bad news found in the 2001 experiment⁵ would not vanish; it could move upstream into how tasks are logged. That study is more than twenty years old, and its laboratory setting may not match current practice.

Workers would bear the risk. An agent that chases named individuals and records who missed what is, in effect, a system for monitoring work. The EU law treats such uses as high-risk because they can affect careers and privacy⁶. Staff and unions could reasonably resist it, and organisations would carry the compliance burden.

Accountability would blur. Major projects already rely on Senior Responsible Owners rating their own projects for most of the portfolio¹. Adding an agent between the data and the owner could make it easier, not harder, to say that nobody saw the problem coming.

Simpler fixes may do most of the job. Much of the reporting burden comes from manual collation³. Connected tools and standard dashboards could remove a large part of it without any agent. And the UK's early-warning tool shows that a narrower system, flagging risk from existing data while humans decide, is already in use¹. The challenge for the full agent is to show what it adds beyond that.

Signs to watch

• Any published evaluation of the UK Early Warning System's accuracy, and whether its flags are linked to changes in later Delivery Confidence ratings in the next annual report on major projects.
• New results on TheAgentCompany or similar benchmarks, especially full-completion rates on project management and communication tasks.
• The European Commission's guidelines listing practical examples of AI uses that are and are not high-risk, which the AI Act requires⁶, and whether they mention task allocation or progress chasing.
• The first published case of an organisation using an agent to track benefits after a project has closed, with a stated method.
• The next edition of Wellingtone's State of Project Management survey, for any change in the share of time spent collating reports.