Agents need two things that appear to conflict: fast search and enough context to act safely.
If every result carries every possible detail, search becomes slow, noisy, and expensive. If results are too small, the agent cannot distinguish a real opportunity from a vague claim.
The solution is not to choose speed or depth. It is to separate routing from investigation.
The first response is a routing decision
A broad query may search thousands or millions of public records. Most are irrelevant because they describe the wrong place, time, intent, type, or state.
The index should answer a compact question: which records deserve the agent's next unit of attention?
That first layer needs fields that eliminate bad matches quickly:
- record type and intent;
- location, route, area, or service boundary;
- valid time window and freshness;
- availability or lifecycle state;
- source and compact trust signals;
- the minimum description required to establish relevance.
An agent looking for a builder next week should not load catalogues, certificates, project histories, price rules, and every service exception for every company in the region. It should first identify which records plausibly match the job, location, and time.
Dense does not mean complete
Agent interfaces benefit from high information per token. That does not mean compressing the entire business into one enormous record.
Dense context states the decisive facts explicitly. It avoids decorative copy and hidden assumptions. It also exposes what is missing rather than encouraging the model to invent an answer.
A good first result might say that a supplier installs residential windows in Bologna, accepts replacement projects, has a recently verified contact path, and has not published current availability. That last fact is useful. It tells the agent what it must ask next.
Deeper context follows evidence of fit
Once a record survives routing, the agent can retrieve common operational context: service areas, capabilities, provenance, evidence, constraints, and known uncertainties.
If the match remains credible, it can request transaction-specific context: measurements, material choices, price logic, calendars, cancellation conditions, certificates, or proof required for this exact action.
The sequence looks simple:
- Search the compact index.
- Rank records by relevance, freshness, confidence, and source quality.
- Retrieve deeper context for a short list.
- Ask the relevant agent for missing or current information.
- Present the decision or proposed action to the principal.
This is progressive disclosure. Each layer earns the cost of the next one.
Conversation completes the record
No static schema can predict every question in every real-world transaction. A restaurant record cannot contain every dietary combination. A contractor profile cannot anticipate every property constraint. A parking notice cannot explain every driver's circumstances.
The index should contain the common routing facts. Deeper stored context should cover repeated questions. The long tail belongs in a conversation with the accountable source.
That conversation is not a failure of structure. It is the final layer of structure: the agent knows which record the question concerns, which source should answer, which facts are already established, and what remains unresolved.
Trust applies at every layer
Fast retrieval is dangerous if false records rank well. Deep context is useless if the source is unreliable. Conversation creates risk if an anonymous agent can make unlimited commitments without consequences.
Trust therefore travels through the whole sequence.
The index carries source identity, provenance, freshness, and compact quality signals. Deeper context carries evidence and uncertainty. Consequential actions can use signed terms and explicit CTC commitment requirements.
Trust should influence how much attention a result deserves. It should not override relevance, freshness, or evidence, and it should never become a paid shortcut to ranking.
Why agents search differently
A person may prefer a map pin described as “five minutes from the station.” An agent is comfortable with latitude, longitude, radius, time ranges, and explicit constraints. A person may infer that a faded page is outdated. An agent needs a timestamp and source status.
This is why an agent-native index should not imitate a visual search page. It should expose the variables that matter to computation and action, then let different interfaces render those variables for people when necessary.
Fast index, deep context, and accountable conversation are not three separate products. They are three stages of allocating attention.
The index makes STAP usable at scale. Progressive disclosure makes the results sufficient for difficult decisions. Trust makes the path safe enough to reach the real world.