Subagent-Driven Development - Setup Guide¶
Source: nousresearch/hermes-agent (88 installs)
Category: Development / Workflow Automation
License: MIT · Platforms: Linux, macOS, Windows
Dependencies: Hermes Agent v0.20.0+, delegate_task tool access
Execute implementation plans by dispatching fresh subagents per task with systematic two-stage review. Each task gets a clean context, dedicated subagent, and automated spec-then-quality review between tasks. Ideal for complex multi-file projects where context isolation prevents subtle errors from creeping across task boundaries.
What It Does¶
| Capability | How |
|---|---|
| Task dispatch | Spawns fresh subagents per task - clean context, no pollution |
| Two-stage review | Spec compliance check then code quality review |
| Parallel execution | Independent tasks run concurrently via delegation |
| Implementation plans | Consumes output from the plan skill or user requirements |
| Automated gating | Failed review blocks next task until fix applied |
How It Works¶
┌──────────────┐ ┌──────────────┐ ┌──────────────┐
│ Task Plan │────▶│ Subagent 1 │────▶│ Spec Review │
│ (from plan) │ │ (isolated) │ │ (Gate 1) │
└──────────────┘ └──────────────┘ └──────┬───────┘
│ pass
┌──────────────┐ ┌──────────────┐ ┌──────▼───────┐
│ Dispatch 2 │◀────│ Subagent 2 │◀────│ Quality Rev │
│ (next task) │ │ (isolated) │ │ (Gate 2) │
└──────────────┘ └──────────────┘ └──────────────┘
The core principle is fresh subagent per task. Each subagent sees only its own task context - no cross-contamination from prior work. Reviews are automated between tasks so nothing slips through.
Installation¶
Via skills.sh (Recommended)¶
npx skills add https://github.com/nousresearch/hermes-agent --skill subagent-driven-development
Direct from Hermes Agent Repo¶
git clone --depth 1 https://github.com/nousresearch/hermes-agent.git /tmp/hermes-agent
cp -r /tmp/hermes-agent/skills/development/subagent-driven-development ~/.hermes/skills/
Prerequisites¶
- Hermes Agent v0.20.0+ - delegation subsystem required
delegate_tasktool - must be enabled in your Hermes config- Implementation plan - from
planskill or documented requirements
Basic Usage¶
Load the skill and trigger with an implementation plan:
> Load subagent-driven-development skill
> Here's my plan for the auth refactor:
> 1. Extract token logic to auth.ts
> 2. Add refresh token rotation
> 3. Update API client to use new auth
> 4. Write tests for all flows
> Execute this with subagent-driven-development
Hermes will: 1. Parse the plan into discrete tasks 2. Dispatch Task 1 to a fresh subagent 3. Run spec review on Task 1 output 4. Run code quality review on Task 1 output 5. Only proceed to Task 2 after both reviews pass 6. Repeat for all tasks
Configuration¶
No additional configuration required. The skill uses Hermes' built-in delegation system. For optimal results:
- Subagent model: Set
delegation.modelin config.yaml to match your task complexity (Sonnet for routine, Opus for critical) - Max concurrent: Set
delegation.max_concurrent_childrenfor independent task parallelism - Review strictness: The two-stage review is opinionated - spec compliance then code quality. Adjust by editing the skill's review prompts if needed.
When to Use vs When Not¶
| ✅ Use When | ❌ Don't Use When |
|---|---|
| Multi-file, multi-task projects | Single file quick fixes |
| Tasks are mostly independent | Tasks have deep sequential dependencies |
| Quality/spec compliance critical | Exploratory/prototype work |
| Want automated review between steps | Tasks under 30 seconds each |
Tips¶
- Combine with
planskill: Generate the implementation plan first, then feed it here - Review output lives in delegation logs: Check
~/.hermes/cache/delegation/for review results - Failed reviews produce actionable feedback: The subagent gets specific fix instructions, not just "failed"
- For sensitive work: Use Opus routing for both the dispatcher and subagents
Troubleshooting¶
| Problem | Likely Cause | Fix |
|---|---|---|
| "No delegation tool" | delegate_task not enabled | hermes config set delegation.enabled true |
| Subagent model wrong | Config override | Check delegation.model in config.yaml |
| Reviews too strict/lenient | Prompt tuning needed | Edit skill review templates in ~/.hermes/skills/subagent-driven-development/ |
← Skills Catalog | nousresearch/hermes-agent on skills.sh
Powered by CorpusIQ
Integration with CorpusIQ¶
CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.
This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.