[FEATURE] Create `Agent` Interface

[mkmeral]

Problem Statement

I would like Strands to extract Agent into an interface. This will allow more complex use cases with multi-agent systems. It can also simplify separate implementations (e.g. MultiAgentBase).

Proposed Solution

Agent Interface Proposal: Extending Strands for Heterogeneous Multi-Agent Systems

Abstract

This proposal introduces an interface-based architecture to replace Strands' current monolithic Agent class, enabling heterogeneous multi-agent systems that combine AI agents, human participants, and external framework integrations within unified workflows. The current architecture limits Strands to homogeneous AI-only systems, preventing real-world applications that require human oversight for regulatory compliance, privacy-conscious processing, or integration with existing enterprise AI toolchains. By abstracting agent capabilities behind a common interface while preserving full backward compatibility, this change transforms Strands from an AI-only platform into a comprehensive orchestration layer for mixed human-AI collaborative systems, addressing critical enterprise needs for compliance, cost optimization, and framework interoperability.

Core Concept

Transform Agent from concrete class to interface - Enable different agent types (AI, human, amnesiac, remote) to participate in multi-agent workflows instead of only LLM-based agents

Couple things:

• Human Agents - Inject human reviewers, approvers, or domain experts into automated workflows for regulatory compliance, quality assurance, and critical decisions
• Amnesiac Agents - Stateless agents that don't retain conversation history, (related issue)
• Remote Agents - Integrate agents from other frameworks (LangChain, AutoGen, CrewAI) or remote systems into Strands orchestration
• ?? Bi-Directional Agents? (not sure)

Interface Definition

class AgentInterface(ABC):
    """Abstract interface for all agent types in Strands."""
    
    @property
    @abstractmethod
    def agent_state(self) -> AgentState: pass
    
    @property
    @abstractmethod
    def agent_id(self) -> str: pass
    
    @property 
    @abstractmethod
    def name(self) -> str: pass
    
    @property
    @abstractmethod
    def agent_type(self) -> str: pass  # "ai", "human", "amnesiac", "remote", etc.
    
    @abstractmethod
    async def invoke_async(self, prompt: Union[str, list[ContentBlock]], **kwargs) -> AgentResult: pass
    
    @abstractmethod
    def __call__(self, prompt: Union[str, list[ContentBlock]], **kwargs) -> AgentResult: pass

# Current Agent becomes AIAgent(AgentInterface) - no breaking changes
Agent = AIAgent  # Backward compatibility alias

Use Case

This would allow customers to intuitively create and manage multi-agent systems including remote agents, agents from other frameworks, or human agents.

Combined Use Case Example

Financial Loan Processing Workflow: Combines AI efficiency, human oversight, privacy compliance, and framework integration

# Mixed agent team for loan processing
loan_processing_team = Swarm([
    # AI agent does initial data analysis
    AIAgent(
        name="data_analyzer", 
        system_prompt="Analyze loan applications and credit history"
    ),
    
    # This is a simple agent that does not update messages between request/responses
    AmnesiacAgent(
        AIAgent(system_prompt="Process low-risk loan approvals"),
        name="routine_processor",
        memory_limit=1
    ),
    
    # Human agent reviews high-risk cases requiring judgment
    HumanAgent(
        name="senior_underwriter",
        description="Human underwriter for high-risk case review"
    ),
    
    # Remote agent runs specialized credit scoring on separate secure system
    RemoteAgent(
        name="credit_scorer",
        endpoint="https://secure-credit-api.company.com",
        auth_token=os.getenv("CREDIT_API_TOKEN")
    ),
    
    # AI agent generates final documentation
    AIAgent(
        name="doc_generator",
        system_prompt="Generate loan approval/denial documentation"
    )
])

# Process loan application with mixed human-AI workflow
result = loan_processing_team("Process loan application ID 12345")

# Workflow automatically:
# 1. AI analyzes application data
# 2. Amnesiac agent handles if low-risk (no data retention)
# 3. Human reviews if high-risk (regulatory compliance) 
# 4. Remote system validates credit score (existing infrastructure)
# 5. AI generates final documentation

Benefits: Regulatory compliance through human oversight, privacy protection via amnesiac processing, integration with existing credit systems, and AI efficiency where appropriate - all orchestrated through a single Strands workflow.

Alternatives Solutions

No response

Additional Context

No response