Build with Hum
Sub-agents
Hum delegates self-contained work to sub-agents that run beside the session on the same workspace. Define your own in .hum/agents/.
Built in
| Kind | Tools |
|---|---|
explore |
Read-only: find things out |
general |
Everything: an independent piece of work, edits included |
The model chooses when to delegate. Several run at once, up to five. A sub-agent never asks you anything.
Your own
One markdown file per agent under .hum/agents/, scanned recursively. Files in .claude/agents/ and .codex/agents/ are read the same way, and ~/.hum/agents/ holds your personal ones. The nearest file wins a name.
---
name: code-reviewer
description: Reviews a change for correctness and style. Use after edits.
tools: read_file, grep, glob
maxTurns: 12
---
You are a reviewer. Report findings with file and line. Never edit.
| Field | Meaning |
|---|---|
name |
The kind the model delegates to. Defaults to the file name |
description |
When to use it. The model reads this to choose |
tools |
Allowlist: read_file, write_file, edit_file, shell, grep, glob, web_fetch, web_search. Claude Code's names are accepted |
disallowedTools |
Denylist |
maxTurns |
Ceiling for one run |
model |
Recorded, not acted on. Every sub-agent runs on the route below |
The body is the agent's role, given to it whole. An agent whose tools can edit is refused while approvals are on.
[model]
agents = "deepseek-v4-flash" # the route every sub-agent runs on; unset = the session's model
Every sub-agent run is on the session transcript, with the file that defined it.