Mathew McKie

What AI agents and organisations can learn from each other

Give five capable people a shared objective and there is no guarantee they will achieve it. They might misunderstand the goal, duplicate work, overlook important information or disagree about who should make a decision. Each person might perform their individual role perfectly well, yet the group still fails.

Now give five capable AI agents the same objective. Some remarkably similar things can happen. As we build systems in which autonomous AI agents divide up work, exchange information and collaborate, we’re encountering problems that organisations have been wrestling with for centuries. How do we give capable individuals a clear purpose, equip them to succeed and organise their efforts into an effective collective? Who should make decisions, how should information flow, and how do we know whether the result is correct?

The resemblance is intriguing. Can knowledge about human organisations help us design better agentic systems? Might designing those systems reveal something about how human organisations actually work? And where does the comparison break down? My conjecture is straightforward: knowledge about organising people should help us ask better questions about organising agents. And the effort to design agents may expose things about human organisations we usually take for granted.

Intelligence is not the same as organisation

We tend to think about organisational capability in terms of the people it employs. Recruit more talented individuals, improve their skills, give them better tools and we might reasonably expect performance to improve. Yet anyone who has worked in a large organisation knows that individual capability is only part of the equation. Highly capable people can spend weeks trying to reach a decision, produce excellent work that contributes little to the overall objective, or duplicate efforts because responsibilities and dependencies are unclear.

Good teams don’t emerge simply because capable people have been put together. J. Richard Hackman’s work identifies six conditions that help: a genuine team, compelling direction, appropriate composition, clear behavioural norms, supportive organisational circumstances and effective coaching.

That aligns with something I’ve repeatedly observed: capable people can achieve remarkably little when their roles are unclear, their objectives vague, their resources inadequate or their colleagues unwilling to cooperate. Effective leadership and incentives matter too, because people need to direct their efforts towards shared objectives rather than simply maximise their individual success.

But there’s an important distinction between understanding what needs to be achieved and being told exactly how to achieve it. Someone who understands the intended outcome can adapt when circumstances change. Someone who knows only the activities they’ve been assigned may continue performing them long after they cease to be useful.

The relevance to AI agents is immediate. An agent can possess considerable reasoning capability yet lack the information, tools, permissions or clarity of purpose needed to contribute effectively. Several agents might have individually sensible instructions that collectively produce conflicting outcomes. These problems aren’t hypothetical. In more than 1,600 multi-agent execution traces, Cemri and colleagues identified 14 failure modes spanning system design, misalignment between agents and inadequate verification.

Human motivation, recognition and fair pay have no straightforward artificial equivalent. But direction, coordination, feedback and resource allocation do have recognisable counterparts in agent design. Intelligence alone does not determine what an individual, or a group, can achieve. The surrounding system matters.

From following instructions to setting direction

As people become more experienced and assume greater responsibility, we increasingly expect them to show initiative rather than wait for instructions. Eventually, some are expected not merely to understand objectives but to help determine what those objectives should be. That is a significant change in the nature of delegation.

Deciding who sets the objective is already a fundamental choice in organising human teams. Ruth Wageman and Colin Fisher distinguish authority over executing tasks, managing work, designing team arrangements and setting direction. The same distinction helps us ask better questions about agents.

Degree of autonomy	Human organisation	Agentic system
Execute	Carry out an assigned task	Follow instructions and execute a workflow
Adapt	Decide how to achieve an agreed outcome	Plan steps, choose tools and change approach
Shape	Identify opportunities and propose new objectives	Suggest priorities or new goals for approval
Direct	Exercise authorised responsibility for strategic direction	Change governing objectives within explicitly delegated authority

An illustrative comparison inspired by research on team authority, not a validated equivalence or reproduction of Wageman and Fisher’s framework.

Whether an agent should ever reach the final level is a governance decision, not a capability threshold. The key distinction is between deciding how to achieve an objective and having authority to change what the objective should be.

Greater autonomy can create value because capable actors can respond to circumstances nobody anticipated. But initiative needs a coherent framework of objectives, responsibilities and decision rights. Otherwise, individually sensible actions can pull an organisation in incompatible directions. Senior leaders sometimes compete for influence, resources and prestige, introducing motivations artificial agents don’t share. Yet independently acting agents may still pursue conflicting objectives if their architecture fails to reconcile them.

The question for designers is therefore not simply how autonomous an agent can be, but how much autonomy it should have, over which decisions, and within what boundaries.

When communication becomes the problem

Even where objectives and responsibilities are clear, cooperation depends on communication. Anyone who has worked across organisational boundaries knows that information can be sent without being understood. People interpret messages differently, overlook qualifications or work from competing versions of what was agreed. Different communication channels offer different opportunities to clarify meaning, challenge assumptions and preserve decisions.

AI agents face related problems, although their communication takes place through software interfaces, messages, shared context and tool outputs rather than human relationships. More agents don’t necessarily make a better team. A 2026 study in Nature Machine Intelligence, involving researchers from Google and MIT, tested 260 agent configurations across six benchmarks, five architectures and three language-model families. Tools, prompts and computational budgets were held constant while the arrangements changed.

Multi-agent coordination helped on some tasks, particularly those that could be divided into parallel activities. On others, especially tasks requiring sequential reasoning, additional agents made performance worse. As individual models became more capable, the benefits of collaboration could also diminish or disappear. Part of the challenge is that exchanging information consumes resources and introduces opportunities for error. Communication between agents can fragment context, increase coordination overhead and allow mistakes to propagate.

Imagine several AI agents working together on a recommendation. One gathers evidence, another analyses it, and a third makes the decision. If information passes along a single chain, one lost message can leave the final agent without something important. But what if the evidence could also reach that agent by another route?

An October 2026 preprint, When Messages Go Wrong, tested exactly this kind of problem. Researchers deliberately disrupted communication between AI agents across more than 29,000 test runs. When communication failed, the arrangement with an alternative route for information still completed 53% of tasks. Two other arrangements, where the necessary information depended on a more vulnerable chain of communication, managed only about 3%.

The difference wasn’t that one arrangement had more intelligent agents. It was that important information could still get through when part of the communication process broke down. The researchers also found that some damaged messages caused failures without triggering an immediate warning. An agent could carry on working without realising that its information was incomplete.

These results come from an early research paper using deliberately introduced faults, so they don’t establish that one organisational structure is always better. But they illustrate something important: how information moves between agents can matter as much as the capabilities of the agents themselves.

This resembles a problem in human organisations. When formal communication channels fail, people sometimes recover through colleagues, informal networks or alternative information sources. But a failure can also travel through several people before anyone realises the original meaning has been lost. The mechanisms differ, but the architectural question is recognisable.

What happens when an important communication fails? Is the information lost, or can the system recover it through another trustworthy path? Communication isn’t simply the exchange of messages. The system must preserve enough information, meaning and context for participants to act effectively.

When collaboration undermines expertise

Surely a team should benefit when one of its members knows the answer. With AI agents, that turns out to be a dangerous assumption. In a 2026 study, Pappu and colleagues examined whether self-organising AI-agent teams could make effective use of their most knowledgeable members. Earlier research had shown how human teams can benefit when they recognise expertise and use it appropriately. The researchers wanted to understand whether artificial teams could achieve similar results.

They frequently could not. Even when told which agent was the expert, teams failed to match that agent’s individual performance across the evaluated benchmarks. On some machine-learning tasks, the losses reached 41.1%. The agents tended to compromise between expert and non-expert opinions rather than giving appropriate weight to superior knowledge. This tendency increased with team size.

It sounds rather like a committee in which consensus becomes more important than finding the right answer. But the result is notable precisely because it differs from what earlier research reports for human teams. The complication is that compromise also improved resistance to adversarial agents. The same behaviour that diluted expertise could also make the team harder to manipulate.

So how much influence should an expert have? More authority might improve accuracy when they’re right, but increase vulnerability when they’re wrong or compromised. Approval can distort performance in another way. Research by Sharma and colleagues found that models trained using human feedback could favour answers aligned with users’ beliefs over more accurate alternatives.

Anyone familiar with a manager who rewards reassuring news but reacts badly to challenge may recognise the behaviour. Yet the causes differ: employees may respond to career incentives or social pressure, while models develop their tendencies through training and evaluation. In both cases, approval can become a poor substitute for correctness. That raises another familiar organisational question: who checks whether the work is actually right?

Who checks the checker?

Imagine asking three AI agents to evaluate whether a company should enter a new market. One researches the opportunity, another builds the financial case, and a third reviews the recommendation. The research agent encounters conflicting estimates of market size but chooses one without preserving the uncertainty. The financial agent uses that estimate to produce an attractive investment case. The reviewer checks the calculations, finds them consistent and approves the recommendation.

Every agent has completed its assigned task, yet an uncertain assumption has become an apparently established fact. The system has provided a review process without necessarily providing meaningful verification. This is the kind of task-verification failure identified in Cemri and colleagues’ research. It also illustrates an old organisational problem: completing a control activity isn’t the same as establishing that the underlying decision is sound.

Organisations have long used separation of duties, independent challenge and decision authority to address similar risks. In computing, the influential Clark–Wilson model drew on commercial control principles to protect information integrity. A more recent agent design makes the distinction tangible. William Waites’ Artificial Organisations describes two reviewing functions: one verifies claims against original sources, while the other assesses the coherence of the argument without seeing those sources.

That is an observed arrangement, not proof that it performs better. But it forces a useful question. Should every reviewer receive all available information, or should access depend on what they’re supposed to establish?

A factual reviewer needs reliable source material. A reviewer assessing an argument’s logic may benefit from considering it independently of those sources. Preserving evidence and distributing evidence are different decisions.

The same distinction applies to authority. Who can approve or reject a result? When should an agent defer to a specialist? How can a reviewer identify errors the original producer is unlikely to recognise? These are familiar organisational design questions. But their answers must be tested in agentic systems, because additional reviews and controls can also introduce cost, delay and complexity.

The objective isn’t to reproduce corporate bureaucracy in software. It’s to establish which arrangements genuinely improve collective performance.

What agents reveal about human organisations

The comparison becomes more uncomfortable when we turn it back on human organisations. When we employ an experienced person, we give them a job description, objectives and access to systems. But we rarely specify everything they need to do to make their role work.

We expect them to interpret ambiguous instructions, resolve misunderstandings, recognise when something makes no sense and exercise judgement about circumstances nobody anticipated. An experienced colleague knows whom to contact when a process breaks down. A manager understands which issues need escalating. A specialist spots assumptions that look reasonable on paper but are questionable in practice. Much of this activity is invisible in formal organisational design.

When engineering AI agents, we have to confront these capabilities more explicitly. How will the system recognise uncertainty, obtain missing information, resolve conflicting instructions or identify situations where it should not proceed? The question exposes something about human organisations: they can depend on capabilities their formal structures barely acknowledge.

This connects to Erik Hollnagel’s work on work-as-imagined and work-as-done: the difference between how activities are expected to happen and how people actually accomplish them. A 2020 study of intravenous medication administration across 16 English hospitals provides a concrete illustration. Using a secondary qualitative analysis of nurses’ accounts, researchers identified how experienced professionals anticipated problems, adapted to unexpected circumstances and exercised judgement to support patient safety.

Some departures from documented procedures reflected reasoned attempts to manage situations the formal processes had not adequately anticipated. The authors weren’t advocating disregard for clinical policies; they were exploring how standardisation and adaptation interact in complex environments. That leaves a question I find much harder to dismiss.

How much of what makes our organisations work is actually designed, and how much depends on people compensating for what we have failed to design?

Not every informal adaptation indicates a weakness. A well-designed organisation may deliberately give people discretion to respond to unforeseen circumstances. But there’s a difference between healthy flexibility and hidden dependence. An organisation might appear to have an effective process when, in reality, an experienced employee is quietly repairing it every day. Everything works until that employee leaves, and a weakness that was always present suddenly becomes visible.

One way to explore this is to ask experienced employees what problems they routinely solve that never appear in their job descriptions, process maps or performance objectives. Then ask what would happen if they stopped solving them. Some answers may reveal valuable expertise and initiative. Others may expose dependencies that formal organisational design has failed to recognise. None of this proves that agent engineering will improve human organisational design. But it may help us see what our formal structures leave out.

What survives the comparison?

The comparison is useful in both directions, but the strength of the evidence is different. Human organisational research offers immediate questions for agent designers. Who defines the goal, and who may change it? What information survives a hand-off? Who can challenge the answer? Experiments already show that architecture affects outcomes, even if they cannot tell us that one arrangement will work best everywhere.

The reverse lesson is more tentative. Designing agents forces us to specify things experienced employees often do without being asked: interpret ambiguity, seek clarification, repair broken hand-offs and recognise when a result makes no sense. It doesn’t prove that agent engineering can redesign human organisations. It might, however, help us recognise what we’ve been taking for granted.

There is also a firm limit to the comparison. Agents don’t share human motivations, relationships or moral responsibilities. Language-model agents also differ from conventional deterministic software because they interpret instructions and adapt to information in ways not fully specified in advance. Similar failures on the surface may have very different causes and remedies.

For executives adopting agentic systems, these aren’t merely engineering choices. They determine what the system is authorised to decide, what evidence it can rely on, when it must escalate and how the business establishes whether the result is worth trusting. Human organisations have spent centuries learning how to establish goals, coordinate specialised individuals, distribute authority, communicate effectively, challenge decisions and create conditions in which capable people can succeed. We’ve also accumulated considerable knowledge about why those arrangements fail.

The practical task is to test the lessons rather than import familiar organisation charts. Agent designers can compare architectures under matched tasks and budgets, then measure how autonomy, decision rights, information paths and verification change accuracy, reliability, time and cost.

Agents aren’t people, so copying human organisations would be a mistake. But ignoring what we’ve learned about organising people would be equally misguided. Borrow the questions organisations have spent centuries trying to answer, and test whether their answers work for agents.