Overview

When a query spans multiple knowledge dimensions, the Orchestrator doesn't process them one by one - it first scans the entire project space, then runs all sub-agents at the same time. This document walks through exactly how that works, using real data from a live three-map run (235,631 points, 3,175 clusters).

Step 1 — SpaceImpressionGraph

The first thing the Orchestrator does is call SpaceImpressionGraph. It takes the user's query and the space_id, and returns a structured summary of every map in the project: names, document counts, cluster counts, and field types.

Example output (abridged):

{
  "reasoning": "The user's query has three distinct components: concepts to study, skills needed, and ongoing research. The space contains exactly three maps that correspond perfectly to these components. Since there are no dependencies between these initial searches, querying all three maps in parallel is the most efficient way to address every part of the request simultaneously.",
  "space_context": {
    "maps": [
      { "name": "Course Concepts Map", "num_documents": 891, "num_clusters": 126 },
      { "name": "Skills Map", "num_documents": 213768, "num_clusters": 2033 },
      { "name": "Research Papers Map", "num_documents": 20972, "num_clusters": 1016 }
    ],
    "space_name": "Edu Knowledge Map"
  }
}

The Orchestrator uses this summary to decide how to route the query — it doesn't assume what maps exist, it discovers them at runtime.

Step 2 — Parallel Execution

Once SpaceImpressionGraph completes, the Orchestrator runs all sub-agents at the same time — one per map. In the Agno Workflow Builder this is set up as three parallel Agent Steps, each with Mantis MCP enabled and a map-specific instruction:

• Course Concepts Agent — Search the Coursera course concepts map to find relevant educational concepts for the user's query.
• Skills Agent — Search the LinkedIn job skills map to find relevant skills required for the user's query.
• Research Papers Agent — Search the ArXiv research papers map to find relevant frontier research for the user's query.

All three run simultaneously with no dependencies between them. The Orchestrator stays in ORCHESTRATING... state until they finish.

Step 3 — Reading the Narrator Pipeline

Open View → Agents → Narrator to see the full execution graph:

KICKOFF → IMPRESSION → TOOL RACE → SYNTHESIS

For a three-map query, the Narrator logs these stages with step counts and latencies:

Stage	Steps	Latency
ORCHESTRATOR	1	—
SPACEIMPRESSIONGRAPH	1	—
IMPRESSIONGRAPH	3 (one per map)	~36s
FILTER	3	~127s
SEARCH	3	~125s
SPATIAL	3	~124s

Each FILTER/SEARCH/SPATIAL row shows 3 steps because all three maps run in parallel at each stage. If you see only 1 step per stage, the Orchestrator routed to a single map — check your agent instructions and confirm Mantis MCP is enabled on each step.

When Parallel Execution Runs

The Orchestrator runs all agents in parallel when there are no dependencies between sub-tasks. A query like "I want to learn machine learning for healthcare — what concepts, skills, and research?" splits cleanly into three independent tasks. A query where one agent's output is needed before another can start (e.g. "find courses that teach the top skills for this job posting") will run sequentially instead.

Known Issues

15-minute canvas stall on large maps (>100k rows) — After processing completes, the map canvas may stay blank for a period. Do not close the tab. Hard refresh after ~15 minutes; points appear after ~60 minutes total on very large datasets.

UI freeze during label generation — The label generation step can cause the browser to stop responding for 2–3 minutes. This recovers automatically with no data loss.