Skip to content

Add delegation-threshold and model-inheritance rules to subagent-orchestration #3

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
assert-not-singularity merged 2 commits into from
Sep 19, 2026
Merged

Add delegation-threshold and model-inheritance rules to subagent-orchestration #3

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a4a8353
Add delegation-threshold and model-inheritance rules to subagent-orch…
assert-not-singularity Sep 19, 2026
d14d788
Make explicit tool restriction the rule for every spawn
assert-not-singularity Sep 19, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 19 additions & 0 deletions plugin/templates/skills/subagent-orchestration/SKILL.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,6 +9,15 @@ description: >

# Subagent & orchestration guidelines

## Don't delegate what a command answers

Before spawning anything, ask whether a grep, one shell command, or reading one file answers it.
If so, do that — a subagent costs its whole tool baseline plus a briefing, to return what a
single call would have. Delegate for a reason you can name: output you won't need in your own
context again, independent work that genuinely runs in parallel, or a fresh perspective on
something you've already formed an opinion about. Checking a claim against a handful of PRs or
issues is a lookup, not a fleet.

## Right-size the model and effort

Default to a **mid-tier model at medium effort**. Escalate to a larger model or higher effort only
Expand All @@ -29,13 +38,23 @@ tokens per agent" — token cost is a real constraint.
Model names change — pick by capability tier, not version. When unsure, start at mid and escalate
only with a stated reason.

Set the model explicitly on every spawn. A spawn with no model inherits its parent's, and a nested
spawn inherits that again — so one top-tier orchestrator silently makes its entire subtree
top-tier, including the mechanical lookups. Most of an over-spent fleet is usually spawned by
other subagents, not by you.

## Restrict each subagent's tools

Every MCP tool schema a subagent loads is billed as a cache **write** on first use (~12.5× the
read price). A file-editing agent that never calls Slack / Jira / cloud / browser tools should not
load their schemas — give it a minimal tool list. This cuts the per-spawn baseline from ~100K+
tokens to a fraction, every spawn.

The default is the failure: a spawn that names no tool list inherits every tool the parent has,
MCP servers included, so restriction has to be explicit on every spawn. Name the few tools the
task needs — a web-research agent needs fetch and search, not a GitHub or Slack MCP — and give it
nothing else.

## Brief each subagent to stand alone

A subagent starts fresh and its final report is **not** shown to the user. So:
Expand Down
Loading