Every Agent Becomes Another Participant in Your Organisation

AI is moving from answering questions to participating in work. The first wave sat beside the organisation: a person asked, a model responded, and the person carried the context, judgment and responsibility into whatever happened next.

The next wave is different. Agents research customers, monitor operations, write and deploy code, update systems, prepare decisions, contact suppliers and trigger workflows. They do not merely produce an answer for a person to interpret. They enter the chain of action.

A company with one hundred employees may soon have hundreds or thousands of machine processes working alongside them. Some will be short-lived. Others will be persistent, specialised and able to act. The relevant organisational chart will no longer contain only people.

“Every agent becomes another participant in your organisation.”

A participant, not just a tool

Calling an agent a participant does not make it a person. It describes its operational role. An agent receives information, forms a view, pursues an objective, coordinates with other actors and changes the environment in which later decisions are made.

Once a machine can recommend which customer to prioritise, change a production system, commit information to a CRM or create work that others depend on, its understanding of the organisation becomes consequential. A partial view is no longer contained inside a chat window. It can become a partial action.

Partial reality scales badly

Human organisations already struggle with fragmented understanding. Different functions see different evidence, use different systems and update their beliefs at different speeds. People compensate through relationships, meetings and memory. Agents inherit the fragmentation but not the informal social machinery that helps humans recognise and repair it.

Each agent is usually given a task, a set of tools and a slice of context. That can be enough to complete a bounded action. It is not enough to ensure that the action remains compatible with the organisation around it.

The failure modes are ordinary rather than spectacular:

• A sales agent uses pricing that leadership has already revised.

• A coding agent implements an architecture that was superseded after a security review.

• A procurement agent negotiates against priorities that operations no longer holds.

• A customer agent repeats a delivery commitment that product now treats as uncertain.

• A monitoring agent sees contradictory evidence but has no route for resolving which view should govern action.

None of these agents needs to be unintelligent. Each can reason well from the world it was given. The organisation becomes incoherent because those worlds are incomplete, stale or mutually incompatible.

“The more agency we give machines, the more important shared organisational state becomes.”

Control is necessary, but it is not enough

The usual response to machine agency is control: permissions, approvals, logs, sandboxes, policies and monitoring. These are essential. They constrain what an agent can do and make its behaviour inspectable.

But control systems tend to answer questions such as: Can this agent call this tool? Did it access this record? Who approved this transaction? They do not necessarily answer: What should the agent currently believe? Which assumptions shaped its objective? What changed after the policy was written? What should happen when two authorised agents reach incompatible conclusions?

An agent can stay within its permissions and still act from the wrong organisational reality. Safe execution is not the same as coherent participation.

The state an agent needs

A prompt describes the immediate task. Shared organisational state supplies the world in which that task makes sense. Before an agent reasons or acts, it needs access to more than relevant documents.

• Position — what the organisation currently holds, including dissent.

• Evidence — why that position is supported and where it came from.

• Uncertainty — what remains provisional, contested or unresolved.

• History — what changed, what was superseded and why.

• Intent — the objectives, constraints, priorities and commitments in force.

• Authority — what the agent may observe, investigate, recommend, draft, approve or do.

• Coordination — who owns related work and which dependencies matter.

• Recourse — where conflict goes when reality cannot be established automatically.

This state must exist outside any one model. Models will change. Agents will be created and retired. Work will move between systems. The organisational world they inherit must persist independently of the participant using it.

Stop reconstructing the organisation on every run

Most agent systems begin with a request, retrieve material, reconstruct the situation, reason, act and then lose much of the working state. The next run starts again. Even when memory is added, it often records what the agent saw or did rather than maintaining the organisation’s current, governed position.

A live organisational model changes that architecture. The world persists between runs. A request arrives into a state that already contains the relevant understanding, evidence, uncertainty, history and authority. The agent can contribute new evidence or a proposed revision, and the state can move without becoming owned by that agent.

This reduces more than retrieval cost. It prevents every machine participant from privately rebuilding the organisation and treating its reconstruction as reality.

AI proposes. Humans ratify.

Shared state should not turn AI into the silent authority on what an organisation believes. New evidence can be extracted by machines. Tensions can be detected. Changes can be proposed. But consequential revisions may still require human judgment, especially where evidence conflicts, objectives compete or responsibility cannot be delegated.

The important design choice is to represent that boundary inside the model. Evidence remains attached. Uncertainty remains visible. Competing views are preserved. Authority determines whether an agent may observe, research, recommend, draft, propose or act. When coordination fails, recourse is explicit rather than improvised.

From independent agents to coordinated participants

People and agents can use different interfaces while operating from the same reality. A person may review a situation in a workspace, inspect the evidence and ratify a revised position. An agent may inherit that revision as structured state, update a plan and monitor whether the expected outcome occurs.

When the evidence changes again, the loop remains connected. The agent does not need to discover that a human changed the plan by finding a new document. The maintained state is the plan’s current meaning, with its evidence, history and authority still attached.

This is what allows agents to become coordinated participants rather than another source of organisational fragmentation. Capability remains distributed. Reality remains shared.

Agency makes shared reality inevitable

The need for shared organisational state grows with every machine actor. A small number of assistants can rely on people to repair missing context. Hundreds of persistent agents cannot. The coordination burden eventually exceeds what meetings, prompt templates and system-by-system integrations can hold together.

Orient is designed for that moment: a maintained organisational reality in which people and agents can reason, coordinate and act, with evidence, uncertainty, history and authority kept explicit.

The question is no longer whether agents will enter organisations. They already are. The question is whether each one will carry its own private version of the company—or participate in a common operating model that can remain coherent as both the organisation and the world change.