Agent Intent Optimisation
Can a human or research agent find a correct, supportable answer to the question that precedes the buyer's decision?
Agent intent optimisation improves the route from a real question to truthful knowledge or an authorised action. It makes four different states legible:
discovery → interpretation → actionability → verification
Search rankings, fetches, citations, and mentions are route signals. None proves that an agent understood the offer, could act, or improved a business outcome.
Changes my mind: Repeated query observations show that structured, sourced answers do not improve correct discovery or qualified use for the target questions.
Model
question → source gap → answer or action route → receipt → qualified use → learning
Job contract
- Valuable outcome: qualified humans and research agents can find, cite, and correctly understand the offer for a real question.
- Trigger and inputs: recurring buyer questions, current answer-surface observations, source pages, query evidence, and claim support.
- Required business skills: customer-language research, information architecture, editorial judgment, source evaluation, and measurement.
- Tool categories and selection: use query research, answer-surface observation, structured-data validation, publishing, and analytics tools that show sources and permit repeatable checks.
- Concrete outputs: query set, evidence gaps, citation-ready answer, discovery-surface receipt, implemented action route where one exists, and observation log.
- Human edge and safe AI: humans approve meaning and claims. AI may cluster questions, inspect structure, and draft from cited evidence; it must not simulate independent demand or citations.
- Proof, cadence, and kill signals: repeat the same queries and surface checks, then observe correct discovery, citation, or qualified action. Review monthly and after material page changes. Stop when content answers no real question, a declared schema or endpoint is unreachable, capability claims outrun implementation, or visibility rises without qualified use.
Classify before optimising
| State | Question | Minimum evidence |
|---|---|---|
| Discovery | Can an agent reach the advertised public surface? | successful fetch of every declared destination |
| Interpretation | Can it identify the provider, purpose, limits, and evidence? | self-contained content with clear authority boundaries |
| Actionability | Can it perform a declared job safely? | reachable invocation binding, inputs, outputs, and receipt |
| Verification | Can it verify identity or the result claimed? | usable verification method or bounded proof |
A knowledge card may pass discovery and interpretation while remaining non-callable. That is an honest capability state, not a defect. A job becomes callable only when its real invocation route and proof exist.
Successful HTTP responses prove availability only. An unreachable schema or advertised service degrades discovery because the public contract contradicts the deployed surface.
Dogfood lesson
In August 2026, a cold-agent inspection of Dreamineering's public discovery
surfaces found that the main files were reachable but one declared schema and
two DID-advertised services returned 404. The agent correctly classified the
site as interpretable knowledge, not a callable capability, and took no action.
The correction removed unsupported promises and added a test that rejects any DID service without an authored public destination. This improves discovery contract integrity. It does not prove citation, invocation, verified identity, or business value.
Practice
Choose one buyer question and inspect the route as a cold agent:
- fetch the public discovery surfaces and every destination they advertise;
- classify each surface as knowledge-only, callable, mixed, or unavailable;
- record identity, authority, evidence, and action gaps separately;
- repair the smallest false or missing connection;
- repeat the same inspection and record action, non-action, or unknown.
Then improve one answer page only when the question and source gap are real. Cite primary evidence, state limits, and route to a useful action that actually exists.
Failure modes
Optimising invented queries creates supply without demand. Chasing citations can weaken the answer. A mention without correct understanding is not success. A descriptive manifest is not a callable capability. A planned endpoint is not a live route. A successful fetch is not beneficiary proof.
Context
- Content Calendar — schedule evidence-backed answer gaps
- Agent-readable conversion — continue from answer to authorised action
- GTM Distribution — place discovery inside a closed route
- Marketing Performance — separate route signals from outcomes
Questions
Next question: Which unanswered buyer question has the clearest evidence of qualified demand and the weakest trustworthy answer?
- What source should a correct answer cite?
- Which observed action would make the citation useful?