Multi-agent coding orchestration: independent contexts + parallel Git worktrees

j3free · reddit · 2026-10-10

The author of open-source extension Pi Herdsman shares an upgraded multi-agent coding workflow. The core idea: keep the orchestrator's context clean — each agent gets its own conversation and execution context, execution history never flows back to the coordinator, and only relevant results are routed back through explicit ownership.

In the new demo, a Manager receives one request to implement two independent features and delegates each to a Lead running in its own Git worktree; each Lead further delegates to an Agent that implements, verifies, commits and reports back. Three components:

Pi: independent conversation, model interactions and execution environment per agent
Herdr: manages physical sessions, processes and workspaces
Herdsman: manages delegation, explicit ownership, result routing, supervision and lifecycle, with nested delegation, continuation, steering and interruption

A key design choice: agents never poll children for results — questions and completions are routed up the ownership hierarchy. Every role has a Markdown Agent Definition (model, instructions, tools, skills, with global/project overrides), so you can experiment with different models without touching orchestration code. The full hierarchy is optional; a single Lead with occasional async agents works fine. Includes a real execution demo and source (GitHub).

Original post →

Multi-agent coding orchestration: independent contexts + parallel Git worktrees

More from coding & agent