Leading Collaborative Centaur Teams

How Leaders Design the Language of Collaboration Between People and Agents.

This is Part 1 of a two-part exploration of Human–AI Teaming Readiness. In this edition, we define the coordination problem that emerges when humans and agents share the same work environment, and why this is becoming a core leadership design challenge. Part 2 will introduce a practical technique that senior leaders can use to deliberately keep collaboration aligned at the enterprise level.

Organisations are in the early stages of introducing new forms of intelligence into environments that were designed exclusively for human collaboration, which means teams are no longer just coordinating across functions and geographies, but also across fundamentally different types of actors, with different modes of perception, speed, memory and agency.

What can leaders do to ensure this works smoothly and successfully?

AI adoption today is still focused on tools. Which copilot should we deploy? Which agents should we orchestrate? Which platforms should we integrate?

But the next question is how coordination actually works when humans and agents share the same operational space. Because coordination is not achieved through tools alone. It depends on shared intent, mutual expectations, clear boundaries of authority, and a common language for what “good” looks like. It depends on trust, not blind trust, but trust that is continually calibrated through feedback and shared understanding. And it depends, above all, on context.

Adding agents into workflows that already struggle with role clarity and decision ownership can exacerbate existing problems. If we ask people to “delegate” to systems whose boundaries of action are poorly defined, some respond by over-relying on (and over-trusting) automation, whilst others respond with defensive scepticism, overriding systems by default. In both cases, coordination degrades because the shared language and context of work is missing or inconsistent.

What emerges is not seamless collaboration, but a new form of friction:

• decisions without provenance,

• actions without clear ownership, and

• learning that fails to accumulate.

This is a coordination design problem that reflects the need for leadership to evolve towards a new craft: the design of the conditions under which humans and machines can act together without constantly pulling the system out of shape.

In mixed-intelligence teams, coordination is no longer something that simply “happens” through informal norms and shared human intuition. It must be deliberately authored. The language of collaboration needs to be designed.

That language is what we mean by context.

This edition builds directly on our recent exploration of world-building as a leadership capability for the agentic era, where we argued that organisations must be designed as coherent worlds of physics (systems), culture (meaning), and geography (experience), and not just as collections of tools and workflows.

Let’s zoom in from the design of the world to the design of collaboration that happens inside it; specifically, how leaders shape the language through which humans and AI work together as centaur teams.

From Tool Use to Teaming Readiness

Collaboration is not a feature of a toolset. It is a property of an environment.

A team does not become a centaur team because it has access to an agent. It becomes one when humans and machines can reliably coordinate their actions around shared intent, shared boundaries and shared meaning. Without that, what looks like collaboration on a process map quickly collapses into a brittle sequence of hand-offs, overrides, and shaky unwritten assumptions.

This is the shift many organisations are now stumbling into without quite realising it.

AI is no longer just something people use. In many settings, it is something that increasingly participates, sensing conditions, drafting actions, monitoring flows, surfacing options, and, in some cases, acting directly in the world. The moment AI participates in work, the question becomes “What does it mean to work together?”

Yet most leadership doctrines, operating models and performance systems were never designed for such a question. They assume human actors with human judgment, human accountability, human learning cycles. Agents enter this landscape as something undefined: sometimes treated as a junior worker, sometimes as a calculator, sometimes as an oracle, sometimes as a risk.

The result is a quiet incoherence in how teams are being asked to relate to their machines.

In some places, delegation runs far ahead of design. Agents are given sweeping autonomy without corresponding clarity on boundaries, escalation, or quality thresholds. In others, distrust freezes collaboration entirely, reducing AI to a glorified drafting aid despite its wider potential. Both patterns look like adoption. Neither look like teaming readiness.

Teaming readiness is something different from tool readiness.

Tool readiness asks:

• Is the technology stable?

• Is it secure?

• Is it integrated?

Teaming readiness asks a different set of questions:

• Do humans and agents share a workable definition of success?

• Is it clear when an agent may act, and when a human must decide?

• Do people trust the system for the right reasons?

• Is learning flowing both from human to machine and from machine back to human?

These are not questions for IT alone. They are questions of leadership design.

Moving from tool use to true teaming requires the deliberate shaping of roles, responsibilities, feedback loops, and the language through which work is coordinated. In other words, through the design of context as a shared operating grammar.

Until that grammar exists, organisations will continue to experience a familiar paradox: impressive local gains from AI deployment, alongside growing systemic fragility in how work actually holds together.

Leadership in mixed-intelligence environments therefore shifts in a subtle but fundamental way. The task is no longer simply to deploy capability. It is to make collaboration itself legible, stable and learnable at the boundary between human and machine.

That is what we mean by Human–AI Teaming Readiness.

What “Context” Means in a Human–AI Team

In human–AI collaboration, context is often treated as a technical concern: prompts, data access, memory, retrieval. These are important foundations. But they are not what makes collaboration work.

In a teaming environment, context is also coordination.

It is the shared frame that allows different kinds of intelligence to act in relationship to one another without constant supervision or repair. It is what tells a human when to trust a system’s output, when to challenge it, and when to override it. It is what tells an agent not just what it can do, but how its actions sit within a wider field of purpose, risk and responsibility.

In world-building terms, this is the moment where physics, culture and geography stop being abstract layers and become the operating language of daily collaboration:

• The System layer provides the physics: rules, data, contracts and constraints that make certain actions possible and others impossible.

• The Culture layer provides the meaning: norms, values, stories and judgments that shape what should happen.

• The Experience layer provides the geography: the interfaces, workflows and spaces through which both humans and agents navigate the world.

Context is the braided fabric of all three.

In this sense, it is not a static asset. It is a living operating language, made up of several intertwined elements:

• Shared intent: a common understanding of what the work is ultimately trying to achieve, beyond the task at hand.

• Boundaries of authority: clarity on when an agent may act autonomously, when it must recommend, and when a human must decide.

• Decision vocabulary: stable definitions of what “good”, “acceptable”, “escalate”, “complete”, or “exception” actually mean in practice.

• Quality thresholds: what level of confidence, evidence, or validation is required before action is taken.

• Risk posture: how much uncertainty the team is willing to tolerate in different contexts.

• Cultural norms of judgment: whether challenge is expected or discouraged, whether speed outweighs precision, whether learning is prioritised over optimisation.

Taken together, these form the grammar of collaboration. Without this grammar, human–AI interaction defaults to two unstable extremes. Either humans over-trust systems, surrendering judgment too early and too broadly. Or they under-trust them, turning agents into little more than sophisticated drafting assistants. In both cases, potential is left unrealised and risk is misunderstood.

The difficulty is that much of this context is normally held tacitly in human teams. It lives in shared experience, informal norms, and unspoken expectations. When agents enter the system, that tacit layer is suddenly exposed. What was once quietly inferred now has to be made explicit if coordination is to hold.

This is why many early centaur team experiments feel awkward at first. The introduction of an agent acts like a mirror, reflecting back the vagueness that already existed within the team.

To become teaming-ready, organisations must therefore do more than supply agents with data and access. They must author the world in which those agents will operate, as physics, as culture, and as navigable experience.

And it is this shared language that leadership is now being asked to design.

Let’s look at a core leadership technique for Human–AI Teaming Readiness we call Align → Bound → Learn, which is a lightweight system for:

• aligning intent,

• designing authority boundaries, and

• ensuring that learning compounds rather than fragments.