When a Hundred Agents Hold a Hundred Versions of the Business

When software moves from answering questions to taking action, business context becomes operational infrastructure.

An organisation acts coherently when the people and systems carrying out a Commitment understand what they are entitled to rely on.

That understanding has always required work. People reconcile conflicting records, ask about exceptions and notice when an instruction no longer fits the situation. They also miss things. Existing automation can carry those mistakes a long way before anyone notices.

Agents extend the range of work that software can initiate. More decisions can happen between human reviews. The assumptions behind those decisions become part of the organisation’s operating conditions.


When an Interpretation Becomes an Action

In a November 1 essay, Yoni Leitersdorf describes a retention agent that mistakes a customer’s move to annual billing for a fall in spending. It initiates an offer even though the apparent decline is explained by the payment schedule.

His example shows how technically successful steps can produce an unsuitable business action. A valid query and a successful API call establish that the machinery operated. The decision also depends on understanding the customer’s situation.

Leitersdorf argues for shared business context as the number of agents grows. I take that argument a step further into Commitment: before an action can create reliance, the organisation needs to establish both its meaning and the authority under which it is taken.

An agent allowed to identify possible churn has a different mandate from one allowed to offer a discount. Access to a customer record does not settle that distinction. Permission to send an offer must be explicit, with limits that apply when the agent acts.


What Needs to Be Shared

Consider the pricing change from the previous essay. Sales needs to know which customers qualify. Finance needs to know when charges become revenue. Support needs to know what the customer was promised.

These are different views of the same relationship. They need to agree where one view creates obligations for another.

Existing software preserves traces of earlier decisions. Some rules reflect obligations that still apply. Others reflect assumptions or workarounds whose circumstances have changed. Reconstructing that logic helps the organisation understand how it arrived here. People still need to establish which assumptions remain valid, which Commitments remain binding and who has authority to revise them.

The shared model must therefore preserve the meaning of the offer, its applicable version and any customer exceptions. It also needs to distinguish observations from assumptions. An old forecast and a current contract should not acquire the same status because both were retrieved in a search.

This does not require one database containing everything or identical access for every participant. Operational records can remain with the systems responsible for them. People and agents need authorised views with clear sources, versions and ownership.

Private preparation and personal learning records can remain private. Organisational coordination needs the evidence people are entitled to use for the decision in front of them.

Nor does shared understanding require universal agreement. A finance objection to a proposed offer may remain unresolved. Keeping that objection visible gives the decision-maker something to address. Removing it to produce a tidy summary would make the model less useful.


Authority Belongs to the Action

The authority question becomes concrete at the boundary between a recommendation and an external effect.

An agent may prepare a proposal from the current offer rules. Another participant may review it. A selected proposal becomes a Commitment only when an authorised undertaking is made. Selection and authorisation can occur together where an explicit mandate permits it.

If the customer’s circumstances fall outside that mandate, the agent needs a defined way to refer the decision to someone who can resolve it. Uncertainty about permission cannot be repaired by generating a more confident explanation.

The resulting record should make it possible to reconstruct which evidence and rules were used, who or what acted and whose authority permitted the action. That is how an organisation can investigate an error and decide which part of the arrangement needs to change.

A useful design also accounts for time. A rule can change between preparation and execution. Before a consequential action, the applicable conditions must still hold or the action must return for review. Copying updated guidance into a shared store does not prove that every agent has adopted it.


Where the Reflective Components Fit

This separation of meaning, evidence and authority has shaped the architecture we are developing at Reflective Labs.

Axioms provides tooling for expressing desired outcomes and constraints in checkable specifications. Such a specification describes what should be true. Evidence is still needed to establish whether a proposed result meets it.

Converge provides the low-level contracts for shared context and governed promotion. A statement admitted into shared state remains a statement supported by particular evidence and rules. Its admission does not establish every claim it makes about the world.

Formation is the planning and orchestration library. It organises attempts and review through formations, bringing different perspectives to a question. A formation can produce a recommendation within its mandate. Binding the organisation requires the relevant authority.

Helms provides a human and operator surface for expressing intent, examining evidence and exercising that authority. Runtime services supply capabilities such as identity, storage and execution around these components.

These responsibilities describe the composition behind the product we are building. Each connected workflow still has to demonstrate that its boundaries hold in practice.


A Model That Can Be Revised

Suppose the retention agent’s mistaken offer reveals that the shared model treated invoice frequency as customer activity. Correcting that definition should improve later decisions. It also raises a separate question about the offer already sent.

A revised belief cannot quietly withdraw a promise. The existing Commitment has to be honoured or explicitly changed by someone with the authority to do so. The record needs to preserve what was known when the original action was taken.

That distinction prevents learning from rewriting history. The organisation can improve its current understanding while remaining answerable for what it did under the earlier one.

This is what I mean by the heart of the organisation: the continuing connection between what people understand, what they have authorised and what others may rely on. As participants and tools change, that connection needs to remain intelligible.

The next essay asks what it means for the organisation to retain custody of it.


Core thesis: Agents need a shared account of the business, authorised access to relevant evidence and explicit mandates for the actions they take.

The idea you can't unsee: Correcting a shared assumption can improve the next decision while leaving an existing promise in force.

Vocabulary shift: "agent integration" → "shared meaning and scoped authority"

Connects to: Article 2 (When Individual Speed Outruns Shared Understanding), Article 4 (What an Organisation Needs to Own), Article 8 of The Organizational Capability (Formations), Report 1 (Proposal, gate, commitment)

Version: 1.2 / 2026-11-07