[agent][efficiency] Orchestrator never fans out sub-agents in parallel — serializes 145-200s apart (defeats parallel/council prompts)

[M3gA-Mind]

Orchestrator never fans out sub-agents in parallel — single-spawns or serializes even when the prompt explicitly asks for parallel/council work

Found via scripted agent-efficiency eval on staging. PARALLEL=0 across the ENTIRE run — every multi-spawn case serialized; several collapsed to a single spawn.

Evidence

• "Spin up a separate researcher for each [Discord/Slack/Circle] IN PARALLEL" → only 1 sub-agent spawned (@15s); ignored the fan-out instruction. Incomplete, empty reply.
• "Convene a model council … multiple independent opinions synthesized" → only 1 spawn (@218s to even start). No council.
• "Fan out and summarize each of my last 4 threads independently" → 3 spawns at 9.8s / 155.2s / 355.5s → inter-spawn gaps 145s and 200s. Strictly SERIAL — each sub-agent spawned only after the prior returned. Never <5s apart.

Ask

Prompts requesting parallel / "separate researcher for each" / council must launch N sub-agents concurrently (spawns within seconds) and reduce them together. This is almost certainly a tinyagents harness issue — check vendor/tinyagents/src/harness/ spawn_parallel_agents / agent_loop concurrency (does it actually run tasks concurrently, or await each sequentially?). This defeats the intent of #4686/#4683. Primary efficiency defect for delegation/council workloads. If the root is in the harness, fix + PR on tinyhumansai/tinyagents (Closes tinyhumansai/openhuman#).