An Agent Development Kit (ADK) for building multi-agent systems with identity management, memory systems, emotion analysis, and MCP/A2A protocol integration
Cognisphere ADK exhibits significant definition quality gaps. Tool names lack consistent verb-noun structure (e.g., 'classify_and_route_query_tool' violates single-responsibility; should be 'classify_query' + 'route_to_agent'). Descriptions are present but generic, most under 100 chars, lacking WHEN/WHY context. Input schemas visible but incomplete: 'tool_context' parameter is underspecified (type 'object' with no properties defined), forcing LLMs to guess its structure. No output schemas documented. Error handling is absent from tool definitions. Tool composition violates single-responsibility: invoke_specialist_agent_tool both loads AND invokes agents. Identity and memory tools lack parameter constraints (e.g., emotion_score 0.0-1.0 stated in description but not in schema; memory_type enum not enforced). A2A tools expose URLs as free-form strings without validation guidance.
Analyzes the emotional content of text.
Classifies the user's query intent and suggests a specialist agent. Also prepares the input for the target agent.
Conecta a um agente externo para processar uma consulta. Connects to an external agent to process a query.
Creates a new identity profile and saves it to persistent storage.
Creates a new memory in the system.
Descobre informações sobre vários agentes A2A. Discovers information about multiple A2A agents.
Dynamically loads and invokes a specialist agent. This version will attempt to use tool_context.actions.transfer_to_agent.
Tool names violate single-responsibility principle. 'classify_and_route_query_tool' combines classification and routing; should split into 'classify_query' and 'route_to_agent'. LLMs cannot infer which sub-action to invoke.
Input parameter 'tool_context' defined as bare 'object' type with no schema properties. LLMs cannot determine what fields to populate. Undocumented dependencies on tool_context.actions.transfer_to_agent and tool_context.session_state.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields. E.g., does 'classify_and_route_query_tool' return agent_id or agent_name? Both? Neither?
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 40 | 2026-07-28+ | v2 |
Lists all available identities.
Recalls memories based on a query and optional filters.
Switches the current context to a specific identity.
Updates an existing identity with new attributes.
Enum constraints missing from schema. 'memory_type' accepts 'explicit, emotional, flashbulb, procedural, liminal' per description but schema shows type 'string' with no enum. 'emotion_type' similarly unconstrained. LLMs will hallucinate invalid values.
A2A tools expose external URLs as free-form string parameters without validation, format, or error recovery guidance. 'connect_to_external_agent' and 'discover_a2a_agents' lack URL validation, timeout handling, and fallback instructions.