Context Is Part of an Agent's Authority

Executive abstract

What an agent can know shapes what it can decide.

Enterprise AI programs often frame context as a retrieval problem: add documents, add records, and let the model sort them out. That framing misses the governance question. The information an agent receives determines what it can infer, combine, disclose, and use to recommend an action. Context is therefore part of an agent's effective authority. Action permissions still matter, but an approval boundary after a broad data dump does not erase the decision power created by the data already supplied. Too little context is not neutral either: omitted constraints, exceptions, or dependencies can make a confident recommendation wrong. The architecture task is to assemble the minimum sufficient, policy-permitted context for a named decision, then retain evidence of what the agent saw and why.

The full-context default

More information is not a neutral control choice.

Sending every available record to a model can increase prompt cost, dilute relevant signal, enlarge the set of facts the system can combine, and make a recommendation difficult to review. It does not follow that every workflow needs an elaborate governance layer. It does mean that a richer route needs a real comparator.

In the released Token-Bleed R5 synthetic experiment, compact governed selection used 96.9% to 97.9% fewer prompt tokens and achieved higher F1 than raw full-context stuffing in its named local configuration. The same experiment also supplied its limitation: against a cheap lexical baseline, governed selection consumed 6.94 times as many prompt tokens on the holdout set, exceeding the preregistered maximum of three. The lexical route scored 0.000 F1 at every catalog size, so governed context won the quality comparison, and the preregistered claim that governance earned its cost still failed. One caveat belongs with that zero: R5 used opaque physical names, where lexical matching has nothing to grip. On an earlier round with a lexically tractable catalog, the same baseline beat governed selection on both quality and cost.

Enterprise Agent Architecture

A context plane belongs beside connection, capability, and evidence.

The distinction is practical. An outage or maintenance exception agent may need weather, load, asset, work-order, and switching-constraint context to recommend whether to keep, reschedule, or escalate a window. It should not receive unrestricted operational data, and it should not obtain dispatch or work-order authority merely because it can assemble a recommendation; that authority must be separately granted, bounded, and evidenced. In CPG, a commercial exception agent may need store, SKU, promotion, inventory, cost, and service context while a named approver retains authority over price or trade terms.

An agent's context is part of its authority.

Decision rule

Start with the least-complex route that could credibly meet the decision.

1. Use a simple filter when it is enough. If it captures relevant records and produces a reviewable decision, do not add machinery for its own sake.
2. Add governed context for a specific reason. Ambiguous field names, business definitions, access policy, lineage, or a required evidence trail may change what a reliable decision needs.
3. Make overhead earn its place. Compare richer context against the cheap baseline on quality, omission risk under routing misses, prompt cost, latency, maintenance work, and human review.

The goal is not more governance. It is a decision-useful context packet with a defensible cost and a clear action boundary.