Agents as Information

AgentsAsInfo.md

Agents as Information

From an information perspective, we can define agents as follows:

1. General Agent — An information-processing system that maintains an internal model of the world and updates it through perception and reasoning to decide actions. It transforms information about the world into behavior.

2. Specialized Agent — A version of the general agent focused on a specific domain or task, using limited world knowledge and task-specific information to act effectively within that scope.

In essence:

A general agent knows about the world; a specialized agent knows how to act within a part of it. Here’s the clean, final form of the three complementary facets—each described consistently, with a single example for every element.

---

Facet 1 — Role / Functional

Describes what role the information plays in the agent’s operation.

• Directive — Constraints that change what is permitted or forbidden. Example: “Never expose PHI in outputs.”

• Instruction — Guidance on how to act; steps, operators, or recipes. Example: “Call flights.search then filter where price < budget.”

• Evaluation — Criteria or tests that judge success or quality. Example: “Block the draft if helpfulness < 3.”

• Feedback — Remediation or advice after an evaluation outcome. Example: “If payment is declined, request a new card.”

• Knowledge — Facts or exemplars that ground reasoning without prescribing or judging. Example: “A Form 10-K is an annual SEC filing.”

---

Facet 2 — Semantic Profile

Describes the internal semantics of the text itself—its intent, force, and level of generality.

• Intent / Direction — Communicative purpose: descriptive | prescriptive | corrective. Example: “Summarize the document in ≤ 150 words.” (prescriptive)

• Normative Strength — Obligation level: consideration → should → must. Example: “You must cite your sources.”

• Scope — Breadth of situations where it holds: global | domain-bounded | condition-bounded | instance-bounded. Example: “For medical summaries, expand acronyms on first use.” (domain-bounded)

• Abstraction Level — Degree of concreteness: principle | template/pattern | example/case. Example: “Prefer concise answers.” (principle)

• Anchoring — Use of domain identifiers or variables: general | semi-specific | task-terms. Example: “Cite Form 10-K Item 7 using section_id.” (task-terms)

---

Facet 3 — Modality

Describes the carrier or representational form in which the information is delivered to the agent.

• Text — Natural-language snippet. Example: “Never expose PHI.”

• Code (Executable) — Procedure expressed in a programming form. Example: def plan_itinerary(origin, dest, date): ...

• Schema (Structured Spec) — Formal I/O or interface definition. Example: OpenAPI schema for flights.search(origin, dest, date).

• Logic / Rule (Constraint DSL) — Declarative expression of conditions or guards. Example: deny if pii_detect(output)

• Tool-Call Spec — Invocation template for an external capability. Example: payments.charge(token, amount) after consent.

• Knowledge Graph — Typed nodes and edges encoding relations. Example: (Form10K) --hasSection--> (Item7: MD&A)

• Vector Index — Embedding representation for retrieval. Example: “10-K interpretation guide” stored as retrievable vectors.