AI Agents Are Rediscovering Organisational Design Problems
In a recent article, I explored why large teams often slow down as they grow. Drawing on the work of Fred Brooks and Robin Dunbar, I argued that adding more people does not simply increase productive capacity; it also increases the coordination required to keep work aligned.
Several people commented about the similar patterns emerging with AI agents and LLM-based systems, and I think it’s an important observation.
The implicit assumption around AI agents often sounds something like: “If one agent helps, then ten agents must help even more.”
But adding participants to a system of work, whether human or artificial, does not just increase output. It also increases communication paths, decision complexity, dependency management and the effort required to keep activity coherent.
The Coordination Problem Doesn’t Disappear With AI
In The Mythical Man-Month, Fred Brooks famously observed:
“Adding manpower to a late software project makes it later.”
His point was not that people are ineffective, it was that coordination carries a cost.
New people require onboarding, context sharing, alignment, communication structures, relationship management and decision-making pathways. Beyond a certain point, the overhead of keeping work coordinated can outweigh the additional productive capacity introduced.
Emerging research into multi-agent systems appears to be rediscovering many of the same constraints. Work by Google Research exploring the scaling behaviour of multi-agent systems found that increasing the number of agents can improve performance for highly parallelisable tasks, but performance often degrades when work requires sequential coordination, shared context or tightly coupled decision-making.
At the same time, conversations around “AI orchestration” are rapidly becoming discussions about communication topology, memory sharing, routing, governance, context management and interaction protocols. Many of these problems resemble the same organisational dynamics that emerge as human systems grow larger and more interconnected.
AI Agents Behave Less Like Tools And More Like Participants
I think this requires a subtle but important shift in thinking. Many organisations still conceptualise AI agents as advanced tools; useful components that can be inserted into workflows to automate or accelerate activity.
But increasingly, agents behave less like static tools and more like participants in systems of work. They receive context, make decisions, exchange information, delegate tasks, trigger actions in other systems, and they rely on shared memory and operational boundaries.
When those things are happening, familiar organisational concerns begin to reappear:
Who owns which decisions?
Where should handoffs occur?
Which agents should communicate directly?
What context should be shared?
How do we avoid duplicated or conflicting actions?
Where does responsibility ultimately sit?
These are fundamental organisational design questions.
AI is Becoming a Stress Test for Organisational Design
I’ve said it before: AI is a stress test for organisational design. It acts as an amplifier for existing dysfunction.
Weak ownership structures become more visible.
Fragmented systems = fragmented agent behaviour = more fragmented systems.
Poorly designed handoffs multiply confusion faster.
Ambiguous decision boundaries generate conflicting outputs.
Disconnected knowledge creates inconsistent reasoning.
In many organisations, humans have historically compensated for these weaknesses informally through conversations, tribal knowledge, workarounds and relationships. But AI agents are much less forgiving.
Which means organisations are increasingly being forced to confront questions they may previously have avoided:
How does work actually flow?
Where are decisions really made?
Who genuinely owns outcomes?
Which interactions are essential, and which are accidental complexity?
These are all valuable questions regardless of the involvement of AI agents.
What Organisations Should Focus on Before Scaling AI Agents
This is the part I think many current AI discussions are missing. The temptation is often to focus on capability first:
Which model should we use?
How many agents should we deploy?
What workflows can we automate?
But before scaling agents, organisations may need to spend more time understanding the systems into which those agents are being introduced.
Start with user needs, not AI capability
Without a clear understanding of who the users are and what needs matter most, AI initiatives can quickly become disconnected from meaningful outcomes.
A user-centric approach provides an orienting constraint for reasoning about where autonomy is useful, where human judgement remains important, and where additional complexity is actually worth introducing. Otherwise there is a risk that different agents, teams and workflows optimise for different local goals. That creates a new kind of fragmentation: highly efficient activity that is poorly aligned to meaningful outcomes.
In previous writing, I explored the idea of “user needs as a north star” for organisational design and AI adoption. The core idea is simple: before scaling capabilities, ensure there is shared clarity about whose needs matter, how value flows, and which outcomes should guide decisions across both human and AI systems.
Without that shared understanding, organisations can end up increasing activity faster than they improve outcomes.
2. Understand the flow of work before scaling agents
Organisations should spend more time mapping and understanding the flow of work before introducing large numbers of agents into operational systems:
where work actually flows
where handoffs occur
where decisions happen
where context gets lost
where ownership is unclear
where teams already struggle to stay aligned
Without that understanding, organisations risk introducing autonomy into workflows they do not yet properly understand. There is a risk that AI simply accelerates fragmentation rather than improving outcomes.
3. Design clear ownership boundaries
One of the most common coordination failures in human systems is unclear ownership.
The same pattern is now appearing in agentic systems:
overlapping responsibilities
duplicated actions
conflicting outputs
circular handoffs
multiple agents attempting to optimise different parts of the same workflow
As these systems become more interconnected, failures also become harder to diagnose. The problem is no longer simply: “What did this agent do?” It increasingly becomes: “What emerged from the interactions between agents?”
Practitioners working on multi-agent systems are already highlighting how unexpected behaviours can emerge through coordination patterns, shared context and feedback loops.
This mirrors many human organisational systems, where outcomes are often shaped less by individual actions than by interactions among participants.
The more agents (human or AI) introduced into a workflow, the more important stewardship and boundary clarity become. Deliberately designed interaction boundaries do not eliminate collaboration, but they do reduce ambiguity around who is responsible for decisions, outcomes and operational integrity.
4. Treat context as infrastructure
One of the more interesting emerging discussions in AI engineering is the growing focus on “context engineering.” Increasingly, many failures in agentic systems are not being attributed to lack of intelligence, but to lack of usable operational context.
AI effectiveness increasingly depends on:
accessible knowledge
structured operational context
clear interfaces
understandable workflows
explicit decision boundaries
clean interaction patterns
In other words, organisational clarity is becoming part of the technical architecture. This has major implications for how organisations think about documentation, platform design, operational modelling and knowledge management.
There is another important implication here: AI agents are not immune to organisational silos. If organisational systems, data and workflows are fragmented, agent behaviour often becomes fragmented too.
Different agents optimise for different local contexts.
Knowledge becomes inconsistent.
Competing interpretations emerge.
Context gets lost at interaction boundaries.
Which means throwing AI at fragmented organisations may simply create faster-moving fragmentation. This is one reason context engineering is becoming so important. The challenge is no longer just connecting models to data. It is creating a coherent operational context that allows humans and AI to reason consistently about work across the organisation.
5. Smaller autonomous units still matter
The temptation with AI is often to centralise capability into large orchestration layers or highly connected operational systems. But the same scaling constraints still exist.
Smaller teams with:
clear domains
bounded ownership
focused workflows
deliberate interfaces
understandable dependencies
are often easier to augment effectively with AI than sprawling, tightly coupled organisational structures. This mirrors many of the same lessons organisations have already learned through software architecture and team design. Autonomy tends to work best when boundaries are understandable.
Final Thought
Many organisations are approaching AI adoption as a capability problem, but the harder challenge may be organisational coherence: designing systems of work where humans and AI can coordinate effectively without overwhelming each other with complexity.
The organisations that navigate this well are likely to be those that understand how work flows, where decisions belong, what outcomes matter, which interactions are essential, and where coordination should end. Because adding AI agents to a fragmented system may simply increase activity faster than it improves effectiveness.
If you want to start raising awareness of how work is flowing within your organisation before adopting AI at scale, take a look at my flow awareness sessions
Sources and further reading
Fred Brooks — The Mythical Man-Month
Google Research — “Towards a science of scaling agent systems”
Snowflake — AI agent orchestration guidance and architecture discussions
Robin Dunbar’s research into social group size and relational limits
Emerging work on context engineering and operational context design
