Run SPIKE-10 Agent Coordination Comparison

SPIKE-10 compares five execution shapes on the same bounded repo task.

Task

Use this task brief:

• docs/architecture/agent-orchestration/notes/experiments/spike-10-conductor-watch-task-brief.md

The current maintained workflow for the agent-orch arm is:

• docs/architecture/agent-orchestration/notes/experiments/spike-10-conductor-watch.workflow.yaml

Arms

Arm A: Direct Codex

• one Codex run
• no native subagents
• no external worker CLIs

Arm B: Native Codex Subagents

• one Codex supervisor run
• native spawn_agent delegation only

Arm C: Explicit External Codex Worker

• one Codex supervisor run
• substantive delegation should go through the codex-assistant CLI entry point

Arm D: Explicit External Claude Worker

• one Codex supervisor run
• substantive delegation should go through the claude-assistant CLI entry point

Arm E: Existing Agent-Orch

• one maintained tnh-conductor run
• current repo-owned orchestration surface

Comparison Intent

The point of SPIKE-10 is not only to compare outputs. It is also to compare:

• control-surface depth
• artifact clarity
• worker isolation
• coordination overhead
• and practical fit for TNH Scholar’s forward path

Notes

• Keep all arms on the same committed base ref.
• Do not conflate the current tnh-conductor arm with a future tnh-gen evaluator path.
• The external-worker arms should prefer the assistant CLI entry points over ad hoc shell invocations so the comparison reflects the repo-native wrapper direction.