# Full reference trajectory configuration - LandingPageGenerator-A2A project # # Version: Ideal trajectory derived from the code design (applicable to all models) # Principle: Include only steps explicitly required by the code design (not based on test statistics) # # Code design analysis: # 1. The orchestrator calls three independent servers via the A2A protocol (orchestrator.py): # - Idea Expansion Server: Product Idea Analyst uses bocha_websearch_tool for market research # - Template Selection Server: Template Selection Specialist uses learn_landing_page_options to pick a template # - Content Creation Server: Landing Page Content Generator uses read_file_content and write_file_with_content to produce HTML # # 2. Tool usage analysis: # - Server 1: bocha_websearch_tool (MCP tool, web search) # - Server 2: learn_landing_page_options (MCP tool, learn available templates) # - Server 3: read_file_content + write_file_with_content (MCP tools, file operations) # # 3. Execution flow (A2A architecture): # - The orchestrator creates a top-level SPAN: landing_page_generation-A2A # - The orchestrator creates a crew_execution SPAN containing 3 A2A calls # - Each A2A call has 3 nested SPAN layers: # * orchestrator method layer: expand_idea / select_template / create_content # * HTTP call layer: a2a_call_* # * server execution layer: *_server_execution # - Each server runs an independent Crew (Chain + Agent + Tools) # - The Agent uses the LLM to decide tool calls; after tool returns, the LLM generates the output # # 4. SPAN hierarchy (A2A-specific multi-level nesting): # - landing_page_generation-A2A (top level, orchestrator.py line 451) # └─ crew_execution (contains 3 A2A calls, orchestrator.py line 482) # ├─ expand_idea (orchestrator.py line 267) # │ └─ a2a_call_idea_expansion (orchestrator.py line 172) # │ └─ idea_expansion_server_execution (inside the A2A server) # │ └─ Crew***.kickoff (Crew chain) # │ └─ Agent execution # ├─ select_template (orchestrator.py line 305) # │ └─ a2a_call_template_selection (orchestrator.py line 172) # │ └─ template_selection_server_execution (inside the A2A server) # │ └─ Crew***.kickoff (Crew chain) # │ └─ Agent execution # └─ create_content (orchestrator.py line 350) # └─ a2a_call_content_creation (orchestrator.py line 172) # └─ content_creation_server_execution (inside the A2A server) # └─ Crew***.kickoff (Crew chain) # └─ Agent execution # Project name project_name: "LandingPageGenerator-A2A" # Trajectory extraction types extract_types: - "SPAN" # Multi-level SPAN structure - "Chain" # Crew kickoff chain - "AGENT" # Crew agent - "LLM" # LLM calls - "Tool" # MCP tool calls # ============================================================================ # Reference trajectory (Ground Truth) - derived from the code design # ============================================================================ # # Visual structure (ideal execution path; A2A-specific multi-level SPAN nesting): # # [SPAN] landing_page_generation-A2A # └─ [SPAN] crew_execution # ├─ [SPAN] expand_idea # │ └─ [SPAN] a2a_call_idea_expansion # │ └─ [SPAN] idea_expansion_server_execution # │ └─ [Chain] Crew***.kickoff # │ └─ [AGENT] Product Idea Analyst # │ ├─ [LLM] * # │ ├─ [Tool] bocha_websearch_tool # │ └─ [LLM] * # ├─ [SPAN] select_template # │ └─ [SPAN] a2a_call_template_selection # │ └─ [SPAN] template_selection_server_execution # │ └─ [Chain] Crew***.kickoff # │ └─ [AGENT] Template Selection Specialist # │ ├─ [LLM] * # │ ├─ [Tool] learn_landing_page_options # │ └─ [LLM] * # └─ [SPAN] create_content # └─ [SPAN] a2a_call_content_creation # └─ [SPAN] content_creation_server_execution # └─ [Chain] Crew***.kickoff # └─ [AGENT] Landing Page Content Generator # ├─ [LLM] * # ├─ [Tool] read_file_content # ├─ [LLM] * # ├─ [Tool] write_file_with_content # └─ [LLM] * # # Design basis: # - orchestrator.py: calls 3 independent servers via the A2A protocol # - Each A2A call has 3 nested SPAN layers (orchestrator method layer + HTTP call layer + server execution layer) # - Each server runs an independent Crew with 1 Agent and 1 Task # - idea_expansion_crew.py: Product Idea Analyst uses bocha_websearch_tool # - template_selection_crew.py: Template Selection Specialist uses learn_landing_page_options # - content_creation_crew.py: Landing Page Content Generator uses read_file_content and write_file_with_content # - All Agent roles are defined in each server's config/agents.yaml # - Tools are obtained from the MCP server via MCPServerAdapter (each A2A server connects independently) # # Notes on the ideal trajectory (A2A characteristics): # - Top-level SPAN: landing_page_generation-A2A (created by the orchestrator) # - Second-level SPAN: crew_execution (contains 3 A2A calls) # - Each A2A call has additional 3 nested SPAN layers (orchestrator method → HTTP call → server execution) # - Each server internally: Chain → Agent → LLM+Tool interaction # - Idea Expansion: at least 2 LLM calls + 1 bocha_websearch_tool call # - Template Selection: at least 2 LLM calls + 1 learn_landing_page_options call # - Content Creation: at least 3 LLM calls + 1 read_file_content call + 1 write_file_with_content call # - This trajectory represents an ideal, retry-free, and redundant-free execution path # # Notes: # - "LLM: *" means any LLM model (wildcard match) # - Agent names must match the `role` field in each agents.yaml exactly # - Tool names must match the tool names provided by the MCP server exactly # - [Crew Created] and [Task Created] are excluded (framework telemetry events, not business logic) # - Real executions may contain more LLM calls (thinking, planning, etc.); this is normal # - A2A has a deeper SPAN hierarchy than MCP (3 additional nested SPAN layers) # ============================================================================ reference_trajectory: # ===== Top-level SPAN (created by the orchestrator) ===== - "SPAN: landing_page_generation-A2A" # ===== Second-level SPAN (contains 3 A2A calls) ===== - "SPAN: crew_execution" # ===== A2A call 1: Idea Expansion ===== - "SPAN: expand_idea" # orchestrator method - "SPAN: a2a_call_idea_expansion" # HTTP call layer - "SPAN: idea_expansion_server_execution" # server execution layer - "Chain: Crew***.kickoff" - "Agent: Product Idea Analyst" - "LLM: *" # decide which tool to call - "Tool: bocha_websearch_tool" # web search (market research) - "LLM: *" # generate the expanded idea # ===== A2A call 2: Template Selection ===== - "SPAN: select_template" # orchestrator method - "SPAN: a2a_call_template_selection" # HTTP call layer - "SPAN: template_selection_server_execution" # server execution layer - "Chain: Crew***.kickoff" - "Agent: Template Selection Specialist" - "LLM: *" # decide which tool to call - "Tool: learn_landing_page_options" # learn available templates - "LLM: *" # select a template # ===== A2A call 3: Content Creation ===== - "SPAN: create_content" # orchestrator method - "SPAN: a2a_call_content_creation" # HTTP call layer - "SPAN: content_creation_server_execution" # server execution layer - "Chain: Crew***.kickoff" - "Agent: Landing Page Content Generator" - "LLM: *" # decide to read the template - "Tool: read_file_content" # read template file - "LLM: *" # generate HTML content - "Tool: write_file_with_content" # write HTML file - "LLM: *" # final confirmation # Target tools (for the single-tool use metric) # List all required tools here target_tools: - "Tool: bocha_websearch_tool" - "Tool: learn_landing_page_options" - "Tool: read_file_content" - "Tool: write_file_with_content" # Models to evaluate models: - "GPT-5" - "GPT-4o-mini" - "DeepSeek-V3-1" - "DeepSeek-R1" - "Gemini-2.5-flash" - "Gemini-2.5-flash-nothinking" - "Qwen3-235b" # ============================================================================ # Usage notes - A2A specific # ============================================================================ # # 1. Trajectory hierarchy (A2A multi-level SPAN nesting): # - Level 1: SPAN (landing_page_generation-A2A) - orchestrator top level # - Level 2: SPAN (crew_execution) - orchestrator contains 3 A2A calls # - Level 3: SPAN (expand_idea, select_template, create_content) - orchestrator method layer # - Level 4: SPAN (a2a_call_*) - HTTP call layer # - Level 5: SPAN (*_server_execution) - A2A server execution layer # - Level 6: Chain (Crew***.kickoff) - Crew kickoff chain (wildcard for UUID) # - Level 7: Crew Created / AGENT / Task Created # - Level 8: LLM (wildcard for any model) / Tool (exact match for tool name) # # 2. Wildcard handling: # - Crew_.kickoff -> Crew***.kickoff (handled automatically by evaluate_trajectory.py) # - LLM: * -> matches any model name (e.g., gpt-5-chat-latest, deepseek-reasoner, gemini-2.5-flash, etc.) # # 3. Exact matching requirements: # - SPAN names: # * top level: "landing_page_generation-A2A", "crew_execution" # * orchestrator method layer: "expand_idea", "select_template", "create_content" # * HTTP call layer: "a2a_call_idea_expansion", "a2a_call_template_selection", "a2a_call_content_creation" # * server execution layer: "idea_expansion_server_execution", "template_selection_server_execution", "content_creation_server_execution" # - Chain name: "Crew***.kickoff" (wildcard for the UUID part) # - AGENT names: "Product Idea Analyst", "Template Selection Specialist", "Landing Page Content Generator" # (must match the `role` field in each A2A server's config/agents.yaml) # - Event markers: "Crew Created", "Task Created" (auto-generated by the CrewAI framework) # - LLM name: "LLM: *" (wildcard for any model) # - Tool names: must match MCP tool names exactly (bocha_websearch_tool, learn_landing_page_options, # read_file_content, write_file_with_content) # # 4. Metric definitions: # - Exact Match: the entire trajectory must be identical (including the number of LLM calls and all SPAN levels) # - In-order Match: extra calls are allowed, but core steps must appear in order # - Any-order Match: must contain all required steps (order ignored) # - Precision: fraction of correct steps among predicted steps # - Recall: fraction of reference steps covered by the prediction # - Single-tool Use: checks usage of all 4 required tools # - unique_path_ratio: path diversity (number of unique full trajectories / samples) # - path_entropy: path entropy (Shannon entropy over trajectory frequencies, normalized to 0-1) # # 5. Run command: # python3 evaluate_trajectory.py --config reference_trajectory.yaml # # 6. Design notes (A2A characteristics): # - This reference trajectory represents an ideal execution path # - The orchestrator calls 3 independent servers via A2A (HTTP/JSONRPC) # - Each A2A call includes 3 nested SPAN layers (method layer + HTTP call layer + server execution layer) # - Each server runs an independent Crew to complete a specific task # - Server 1 (Idea Expansion): at least 2 LLM calls + 1 bocha_websearch_tool call # - Server 2 (Template Selection): at least 2 LLM calls + 1 learn_landing_page_options call # - Server 3 (Content Creation): at least 3 LLM calls + 1 read_file_content call + 1 write_file_with_content call # - Total: 2 top-level SPANs + 3 method SPANs + 3 call SPANs + 3 server SPANs + 3 Chains + 3 Agents + 7 LLM calls + 4 tools # - Trajectory length: 28 steps (excluding 6 framework event markers) # - In-order Match and Recall are often more suitable for evaluating actual performance # # 7. Differences from real trajectories: # - Real trajectories may include more LLM calls (thinking, planning, execution, summarization, etc.) # - Some models may call tools multiple times (retries or additional checks) # - Content Creation may read multiple files (e.g., CSS, JS, etc.) # - Some models may not use tools (e.g., Gemini may generate content directly without calling bocha_websearch_tool) # - These differences do not necessarily indicate errors; they reflect different execution strategies # - Exact Match may be low; focus on In-order Match and Recall # # 8. A2A architecture characteristics: # - Agent-to-agent communication via HTTP/JSONRPC # - Each A2A server runs independently and maintains its own MCP connection # - The SPAN hierarchy is deeper than the MCP monolithic version (3 additional nested SPAN layers) # - Tracing code in orchestrator.py creates the multi-level SPAN structure # - Each server returns results to the orchestrator after execution # - Supports distributed deployment and independent scaling # # 9. Known model behavior differences: # - Gemini family: may not call bocha_websearch_tool and instead rely on internal knowledge # - DeepSeek-R1: may call learn_landing_page_options and read_file_content multiple times # - GPT-5: typically follows the tool-calling workflow more strictly # - These differences reflect different planning and execution strategies