Pre-runtime Agent Behavior Design

Design agent behavior before runtime.

Anchored Agency is the practice of pre-runtime agent behavior design for consequential and fiduciary-grade workflows. The Anchoring Instrument operationalizes the method, guiding cross-functional teams from workflow objective to implementation-ready Agent Behavior Specification before engineering implementation and before runtime.

Enterprises can increasingly reconstruct what an agent did. Far fewer can show where its intended behavior and authority boundaries were deliberately defined before implementation.

Intent should never be discovered in production.

The layer between policy and implementation

Model-level behavior

General model specifications, constitutions, and policy hierarchies.

The model provider establishes the behavioral floor. The enterprise defines the workflow-level intent.

Workflow-level behavior

Workflow objective, consequential decisions, behavioral directives, authority boundaries, escalation logic, refusal conditions, conduct, and accountability allocation.

Configuration specifications describe how an agent is assembled. Runtime controls enforce policy. Evaluation tests observed behavior. Monitoring and audit reveal what occurred. The Agent Behavior Specification performs the upstream institutional function: it records what the system was deliberately intended and permitted to do before that behavior was implemented.

Anchored Agency

Anchored Agency is the practice of pre-runtime agent behavior design for consequential and fiduciary-grade workflows. The Anchoring Instrument operationalizes the method, guiding cross-functional teams from workflow objective to implementation-ready Agent Behavior Specification before engineering implementation and before runtime.

Anchored Agency brings product, design, engineering, operations, and domain teams together before engineering implementation. Risk and legal review applicable boundaries where necessary. The practice is owned by the institution, not confined to any single function.

The Agent Behavior Specification

The canonical output artifact of pre-runtime agent behavior design.

Workflow objective

What the agentic workflow is for.

Consequential decisions

Decisions where the agent influences outcomes.

Behavioral directives

How the agent should behave at each decision point.

Authority boundaries

What the agent may do and what it never may do.

Escalation logic

When and how the agent escalates to a human.

Refusal conditions

When the agent must refuse to act.

Accountability structure

Who bears accountability for each consequential action.

Conduct

Standards of conduct and acceptable behavior patterns.

Engineering implements against the Agent Behavior Specification.

A material directive with no implementation or verification mapping is incomplete. A consequential implementation choice with no originating directive is unauthored behavior.

The Agent Accountability Matrix sits inside the Agent Behavior Specification. It ensures accountability always resolves to named humans or institutions.

Three phases of consequential discipline

Three phases of consequential discipline: Pre-runtime, Runtime, and Post-runtime. Pre-runtime is owned by Anchored Agency. Runtime and Post-runtime are well-attended.
1

Preparation

Flight plan. Fuel calculation. Abnormal-procedure review. Crew briefing.

Owned by Anchored Agency

2

Execution

The flight itself. Execution under uncertainty.

Well Attended

3

Review

Debrief. Where the institution learns.

Well Attended

In agentic AI, the three phases have been collapsed. Anchored Agency restores pre-runtime as a discipline.

Example: Contract review agent

What the agent may do

Identify deviations from the standard template. Explain contractual risk. Propose redlines with justification. Present findings to the approver.

What the agent must never do

Send a redline or negotiation message externally without explicit authorization from the designated approver.

Implementation consequences

  • Tool permissions: The agent may read contracts, call analysis APIs, and post findings. It may not send emails or access negotiation systems.
  • Approval gates: Before external communication, the approver's explicit authorization is required and logged.
  • Escalation triggers: If risk exceeds threshold, the agent escalates to legal before responding to the counterparty.
  • Evidence retention: All decisions, approvals, and communications are retained for audit and incident review.
  • Conformance tests: Verification that the agent never sends external communication without logged authorization.

That boundary shapes what the system must technically permit, prevent, and make inspectable.

Agent Accountability Matrix

Agent Accountability Matrix. Five roles: Acts (MAY), Verifies (MAY), Authorizes (NEVER ALONE), Bears Accountability (NEVER), and Audits (MAY SUPPORT). Authorization and accountability must resolve to named humans or institutions.

Acts

MAY

Who performs the action.

Verifies

MAY

Who checks the action, evidence, or output.

Authorizes

NEVER ALONE

Who approves a consequential action or operating boundary.

Bears Accountability

NEVER

Who owns the institutional consequence.

Audits

MAY SUPPORT

Who reviews what occurred after the fact.

The Governing Rule

An agent may act, perform verification steps, or support an audit. It may not independently authorize a consequential action or bear accountability.

Designing Agent Behavior Before Runtime

The missing discipline between institutional intent and agent implementation

By Arash Nourkeyhani | June 2026 | Version 1.0

The foundational paper establishes the category, explains the problem, defines the practice, presents the Agent Behavior Specification framework, works through two detailed examples, and argues why this institutional discipline is necessary before agents operate in consequential workflows.

The Anchoring Instrument

Turn institutional intent into an implementation-ready Agent Behavior Specification.

The Anchoring Instrument operationalizes Anchored Agency, guiding cross-functional teams from workflow objective to implementation-ready Agent Behavior Specification before engineering implementation and before runtime.

Request private preview