Dynamic Behaviors

What are Dynamic Behaviors?

Dynamic behaviors are a unified system that enables conversational agents to manage knowledge, intuition, side-effects, actions, and data integration. Effectively, it allows the selection agent to examine available options to modify its own local context graph topology and cause side-effects. This system enables more natural, context-aware responses that enhance the user experience while still maintaining necessary guardrails.

How Dynamic Behaviors Work

Dynamic behaviors operate on a spectrum from strict compliance requirements (redlining) to flexible user experience enhancements. Here's how the system processes them:

1. Conversational Triggers

• Triggers are patterns or keywords in user messages that may activate a specific behavior

• Triggers are not meant to be exact or close matches - they function as relative ranking mechanisms only

• Triggers can be:

  • Completely associative tags (e.g., "exercise," "anger," "protein")

  • Specific identifiers mentioning particular items (e.g., a specific drug or food like "chicken")

  • Thematic/conceptual: Broader topics or semantic areas (better for product experience)

  • Exact matches: Precise phrases that must appear (ideal for safety/compliance cases)

2. Candidate Evaluation and Ranking

• The system presents candidate instructions to the behavior selection agent

• The selection agent conditions against interaction logs and user model

• Candidates are ranked by relevance to the current context

• Only the relative ranking matters, not absolute scores

• Think of this as "the agent moving around the conversation with relevant content to optionally grab or ignore"

3. Selection Decision

• The behavior selection agent has three options:

  • Continue using the previous dynamic behavior (if any) to maintain thread consistency

  • Select a new candidate behavior that better fits the current context

  • Select nothing if no behavior is appropriate

4. Context Graph Integration

• If a behavior is selected, it is injected into the action guidelines of the current context graph state

• This modifies the problem topology, directly influencing agent behavior

• The selected behavior becomes part of the agent's decision-making framework, but doesn't guarantee 100% enactment

• The behavior merges with the local context graph state to create a more constrained, guided, or contextually rich environment

• Every time a dynamic behavior is selected, the context graph is modified

• The dynamic behavior gives higher resolution and scales much larger to ~5 million characters (without side-effects, with side effects this can scale another order of magnitude larger)

• The behavior can:

  • Enrich the response (for user experience)

  • Constrain the response (for safety/compliance)

  • Provide additional guidance

  • Add new tool exposure

  • Trigger hand-off to external systems

  • Add new exit conditions

  • Enable reflection and self-modification

5. Contextual Implementation

• The agent autonomously decides how to implement the behavior in a natural way

• The behavior is integrated into the response rather than robotically inserted

Implementation Best Practices

Design Along the Instruction Flexibility Spectrum

Instructions can range widely in their prescriptiveness:

Open-Ended Guidance

• Allows significant agent discretion

• Provides general direction without strict requirements

• Creates a knowledge-enriched environment for the agent

• Best for creative, exploratory, or coaching conversations

• Example: "You have access to these nutrition resources. Consider their relevance to the user's goals."

Structured Protocols

• Provides clear, specific guidelines

• Balances direction with some contextual adaptation

• Ensures consistency while allowing natural conversation

• Example: "When discussing exercise plans, ask about previous injury history before making recommendations."

Strict Instructions

• Enforces precise behavior patterns

• Minimizes agent discretion for safety-critical scenarios

• Essential for regulatory compliance or high-risk situations

• Example: "If the user mentions suicidal thoughts, immediately provide crisis resources and follow safety protocol X."

Separate Redlining from Product Experience

Split your dynamic behaviors into two categories:

Redlining (Safety/Compliance)

• Purpose: Protect against medical emergencies, legal issues, or user harm

• Trigger Design: Exact patterns that must be detected

• Instructions: Can be more directive and specific

• Testing: Subject to unit tests that must pass for every release

• Example: Suicide prevention protocol, legal disclaimers

Product Experience

• Purpose: Enhance user experience, provide relevant content

• Trigger Design: Broader thematic matches (e.g., "exercise", "protein", "recipes")

• Instructions: Provide options rather than mandates (e.g., "You have access to these resources, surface them if relevant")

• Testing: Measured via metrics and user experience evaluation

• Example: Suggesting relevant meal plans when user discusses nutrition

Measurement Approaches

For Redlining

• Implement unit tests

• Every critical safety case must be tested and pass before release

• 100% compliance expected

For Product Experience

• Develop metrics that focus on:

  • Is content being served appropriately?

  • Is content relevant to the conversation?

  • Is content repetitive?

  • Overall user experience quality

• Monitor trends over time (e.g., 30-day metrics)

• Audit samples where expected behaviors didn't trigger

• Focus on anomaly detection rather than expecting 100% triggering

Example Implementation

Poor Implementation (Too Rigid)

Trigger: "protein"
Instruction: "You MUST suggest the following three protein shakes EVERY time a user mentions protein..."

Better Implementation (Natural)

Trigger: ["protein", "nutrition", "workout", "exercise", "diet"]
Instruction: "You are an in-app assistant with access to 10 high-protein recipe deep links. Each has a brief description. You may surface these when relevant to the conversation about nutrition or fitness. Use your judgment on how many to suggest (max 3) and integrate naturally into your response. Don't repeat recommendations."

Agent Autonomy Spectrum

The design of triggers and instructions directly impacts agent autonomy:

High Autonomy Configuration

• Uses vaguer triggers combined with open context

• Grants the agent greater freedom to determine behavior based on user model and interaction context

• Functions like an associative knowledge cluster available to the agent

• Agent can selectively draw from this knowledge as the conversation evolves

• Best for creative coaching, exploratory discussions, and personalized experiences

Limited Autonomy Configuration

• Uses stricter triggers paired with precise instructions

• Effectively simulates protocol overrides in critical situations

• Limits agent discretion and enforces specific response patterns

• Necessary for regulated industries, safety-critical information, and compliance requirements

Strategic Implementation

• Most effective systems use a thoughtful mix across the autonomy spectrum

• Critical areas employ strict triggers with precise instructions

• General conversation areas use broader triggers with flexible guidance

• Creates a balanced system that's both compliant and conversationally natural

Common Pitfalls

1. Over-engineering: Creating overly specific triggers that rarely match

2. Robotic responses: Making behaviors too directive, leading to unnatural interactions

3. Expecting 100% triggering: Product enhancements should be contextual, not forced

4. Mixed responsibilities: Having the same team design both safety and product behaviors

5. Inconsistent testing: Using the wrong evaluation approach for different behavior types

6. Misunderstanding behavior integration: Expecting behaviors to be rigidly enacted rather than merged with the context graph state

7. Ignoring the selection process: Not accounting for the three options (continue previous, select new, select nothing) in behavior design