A Source of Truth Needs a Write Path

Agents need durable context. A worker that wakes up with only a ticket and a repository has to reconstruct the product, the constraints, and the decisions that shaped the code. It will fill missing context with plausible assumptions. Another worker will fill the same gaps differently. The result looks like autonomy from a distance and contradiction up close.

A source of truth changes that. Product intent, architectural boundaries, accepted decisions, operating standards, and current system state become inputs the factory can retrieve instead of folklore a person must repeat. The work survives a model swap, a fresh session, and the departure of whoever used to hold the explanation in their head.

That is real leverage. It is also only the read side of the system. Once agents can act, they do not merely consume truth. They create candidate facts: a decision was made, a requirement changed, a deployment occurred, a check failed, a production behavior appeared. The factory has to decide which of those claims may become shared state.

A folder full of documents is not authoritative because everybody can open it. Authority comes from a write contract. Which system owns each fact? Who or what may change it? What evidence must accompany the change? What happens when two writers disagree? If those answers live in convention, the source of truth is just the latest file somebody touched.

The contract should be mechanical. An agent can propose a requirement update, but the product owner accepts the changed destination. A builder can report a completed change, but an independent verifier supplies the acceptance evidence. A production observer can record behavior, but it cannot quietly rewrite the standard that declares the behavior acceptable. Different facts need different authorities.

This is where human judgment moves higher. People choose the destination, define the standards, and own consequences that cannot be delegated. They should not spend their day copying decisions between tickets and documents. That transcription is repeatable work, and repeatable work belongs to agents once the write contract is explicit.

Compression is unavoidable. No worker should reread the entire history before changing one component. But a concise context packet becomes dangerous when it loses the path back to the decision, evidence, and scope that produced it. A summary can outlive the condition that made it true. Then an old workaround silently becomes architecture.

Every durable claim needs provenance: the triggering request, the authority that accepted it, the artifact it governs, the evidence that supports it, and the conditions that invalidate it. The agent can receive the compact version first and retrieve the chain when consequence or ambiguity demands it. Context stays small without becoming unaccountable.

Provenance also makes correction possible. When production disproves an assumption, the factory can find every specification, check, and route derived from it. Without that link, somebody edits a page and hopes the old belief is not still embedded in prompts, tests, runbooks, and pending work.

Two agents will eventually produce incompatible claims. One sees the documented requirement. Another sees production behavior. A third finds a code path that contradicts both. Asking a model to merge the accounts into one smooth paragraph destroys the most valuable signal in the system: the disagreement itself.

The factory should preserve the conflict, identify the authorities involved, and route the decision according to consequence. Some conflicts resolve through fresh observation. Some require a specification owner. Some expose a broken control and must block release. The system becomes trustworthy by refusing to manufacture certainty where none exists.

A source of truth is therefore not a single voice. It is a governed process for deciding which claims hold, for what scope, and until what evidence changes them. The record is singular because the decision procedure is shared, not because ambiguity has been deleted from the page.

The final test is behavioral. If a declared boundary does not constrain tools, it is advice. If an accepted requirement does not generate checks, it is prose. If a production failure does not alter the next plan, it is trivia. Operational truth must change permissions, routes, budgets, gates, or acceptance conditions.

Build one governed record, then connect it to execution. Let agents maintain the repeatable links, collect evidence, detect conflicts, and prepare proposed updates. Keep authority explicit. Require independent proof for claims that unlock consequential action. Retain the history so the next worker inherits the reason, not just the sentence.

Our prediction is that organizations will not differentiate themselves by stuffing more documents into a model window. They will win by turning shared state into enforceable machinery. The model will change. The workers will multiply. The source of truth will survive only if every write has an owner, every claim has a trail, and every accepted fact changes the system that acts on it.