PRD-027: External Agent Interoperability

Status: Active Version: 1.0 Date: 2026-01-02 Satisfies: ADR-038, ADR-037

Purpose

Define requirements for enabling SEA™ agents to communicate and collaborate with external AI agents using the A2A (Agent-to-Agent) protocol, and to orchestrate internal workflows using Microsoft Semantic Kernel.

Problem Statement

SEA™ agents currently operate in isolation within the SEA™ ecosystem. They cannot:

• Receive tasks from external AI agents
• Delegate sub-tasks to specialized external agents
• Expose their capabilities for discovery
• Maintain conversational memory across agent interactions

This limits SEA™’s ability to participate in the emerging multi-agent ecosystem where agents from different frameworks (LangGraph, CrewAI, Google ADK, etc.) collaborate on complex tasks.

Goals

1. Multi-Agent Collaboration: Enable SEA™ agents to collaborate with external AI agents
2. Capability Discovery: Expose SEA™ agent capabilities via standardized Agent Cards
3. Governed Orchestration: Ensure all agent interactions follow SEA™ governance rules
4. Memory Integration: Provide semantic memory for context retention across interactions

Non-Goals

• Real-time audio/video agent communication (out of scope)
• Agent marketplace or registry (post-MVP)
• Custom A2A protocol extensions (follow standard only)

Requirements

REQ-070: Agent Card Discovery

Priority: P0 (MVP) Type: Functional

The system SHALL serve an A2A Agent Card at /.well-known/agent.json that:

• Describes agent identity (name, version, description)
• Lists agent capabilities and skills
• Specifies supported authentication schemes
• Declares supported features (streaming, push notifications)

Acceptance Criteria:

• Agent Card validates against A2A schema
• Skills derived from Semantic Kernel plugins
• Authentication schemes include OAuth 2.0 bearer tokens

REQ-071: Task Reception

Priority: P0 (MVP) Type: Functional

The system SHALL accept tasks from external A2A clients:

• Receive SendMessage requests with task context
• Create internal Task objects with unique IDs
• Route tasks to appropriate Semantic Kernel plugins
• Return task status and artifacts

Acceptance Criteria:

• POST /tasks creates new task
• Task lifecycle states match A2A spec
• Task ID format: sea-task-{ulid}

REQ-072: Message Exchange

Priority: P0 (MVP) Type: Functional

The system SHALL support multi-turn message exchange within tasks:

• Accept messages with Parts (TextPart, FilePart, DataPart)
• Maintain conversation history within task context
• Support input-required state for human-in-the-loop

Acceptance Criteria:

• POST /tasks/{id}/messages adds message to task
• Messages support all A2A Part types
• Conversation history retrievable via GET /tasks/{id}/messages

REQ-073: Governed Task Execution

Priority: P0 (MVP) Type: Security

The system SHALL enforce governance rules for all task executions:

• Semantic Kernel HITL filter for mutating operations
• Audit filter for all executions (success and failure)
• Authority boundary enforcement per SDS-031

Acceptance Criteria:

• Mutating operations require HITL approval
• All executions logged to audit trail
• External tasks cannot bypass SDS-031 rules

REQ-074: Task Delegation

Priority: P1 (Post-MVP) Type: Functional

The system SHALL delegate sub-tasks to external A2A-compliant agents:

• Discover external agents via Agent Card
• Send tasks with appropriate context
• Handle async responses and streaming updates

Acceptance Criteria:

• A2A client can fetch external Agent Cards
• Task delegation with configurable timeout
• Streaming updates consumed via SSE

REQ-075: Streaming Updates

Priority: P0 (MVP) Type: Functional

The system SHALL provide real-time task updates via streaming:

• Server-Sent Events (SSE) for streaming responses
• Task status updates (state changes)
• Incremental artifact delivery

Acceptance Criteria:

• SSE endpoint at /tasks/{id}/stream
• Status events within 100ms of state change
• Artifacts streamed as available

REQ-076: Push Notifications

Priority: P1 (Post-MVP) Type: Functional

The system SHALL support push notifications for long-running tasks:

• Webhook registration for task updates
• Retry with exponential backoff on failure
• HMAC signature for webhook validation

Acceptance Criteria:

• Webhook registration via POST /webhooks
• Retry up to 5 times with exponential backoff
• HMAC-SHA256 signature in X-A2A-Signature header

REQ-077: Semantic Memory

Priority: P0 (MVP) Type: Functional

The system SHALL maintain semantic memory for agent interactions:

• Vector storage via pgvector
• Memory scoped to task contexts
• Semantic search for context retrieval

Acceptance Criteria:

• Memory adapter integrates with Semantic Kernel
• Vector embeddings stored in pgvector
• Context retrieval within 50ms

REQ-078: Plugin Capability Mapping

Priority: P0 (MVP) Type: Functional

The system SHALL map Semantic Kernel plugins to A2A Agent Card skills:

• Plugin functions become skills
• Plugin metadata drives skill descriptions
• Capability flags derived from plugin attributes

Acceptance Criteria:

• Agent Card skills auto-generated from plugins
• Skill IDs match plugin function names
• Skill descriptions from plugin docstrings

User Stories

US-070: External Agent Collaboration

As an external AI agent (LangGraph, CrewAI, etc.) I want to discover and send tasks to SEA™ agents So that I can leverage SEA™’s semantic engineering capabilities

Acceptance Criteria:

• Given I have the SEA™ agent URL
• When I fetch /.well-known/agent.json
• Then I receive a valid A2A Agent Card
• And I can send tasks using the declared endpoint

US-071: Governed Execution

As a Security Officer I want all external agent tasks to pass through governance filters So that external agents cannot bypass SEA™’s security rules

Acceptance Criteria:

• Given an external agent sends a mutating task
• When the task is processed
• Then HITL approval is required
• And the execution is logged to audit trail

US-072: Context-Aware Responses

As an external AI agent I want SEA™ agents to remember conversation context So that I can have multi-turn interactions

Acceptance Criteria:

• Given I send multiple messages to a task
• When I reference previous context
• Then SEA™ agent uses memory to understand references
• And responses are contextually appropriate

Dependencies

Success Metrics

Risks

References

• ADR-038: A2A Protocol & Semantic Kernel Integration
• A2A Protocol Specification
• Semantic Kernel Documentation