agent-offload launches another coding agent and blocks until that agent
finishes. Use it when your main agent wants to delegate implementation work
while keeping the current conversation in control of review, verification, and
final reporting.
Why?
Sometimes you already have a thorough Markdown plan, and implementing it with a heavy, slow model is overkill. The host agent should be able to hand that work to a cheaper and faster model directly, without user manually starting a new Codex, OpenCode, or other agent session.
Claude Code can delegate to subagents, but that keeps delegation inside Claude's own harness and model choices. Other agent harnesses have their own delegation mechanisms, assumptions, and model support. For example, you might be inside Claude Code with a plan ready to implement, but want Codex Spark or a DeepSeek V4 backed agent to do the edit. The host agent should be able to start that run, pass the plan, wait for completion, and review the result without asking the user to open and manage a separate session.
agent-offload makes offloading harness agnostic by using process execution as
the boundary. The host agent runs one blocking command, the delegated task opens
in a new tmux pane or runs headlessly, and the host agent waits until the task
finishes. Then the host agent can inspect the diff, run checks, and continue the
review in the original conversation.
What it does
- Launches configured agent profiles in tmux or in headless mode
- Sends prompts as an argument, stdin, or a prompt-file argument
- Adds completion-file instructions to tmux delegated prompts
- Waits until the delegated agent exits or publishes
done.md - Detects if the tmux pane exits before completion
- Kills the delegated pane after completion
- Marks Cursor Agent workspaces as trusted before launching them
- Supports per-profile environment variables, including forwarding from the host environment
- Installs its provider-agnostic skill bundle with
agent-offload install-skill
Quick start
1. Install
# Shell script, macOS/Linux
|
# Homebrew
# Cargo from source
2. Create a config
agent-offload reads:
~/.config/agent-offload/config.yaml
See Configuration for an example.
3. Install the skill
This writes:
~/.claude/skills/agent-offload/SKILL.md
~/.config/opencode/skills/agent-offload/SKILL.md
~/.codex/skills/agent-offload/SKILL.md
~/.pi/agent/skills/agent-offload/SKILL.md
Then use /agent-offload from your host agent UI, or call the CLI directly.
4. Delegate work
From Claude Code, invoke the installed skill:
/agent-offload implement the change in history/plan.md
The host agent loads the skill, chooses a configured profile, runs
agent-offload, waits for the delegated agent's summary, then reviews and
reports the result in the current conversation.
How it works
Each run gets a directory under:
~/.local/state/agent-offload/runs/
The run directory contains:
| File | Purpose |
|---|---|
prompt.md |
The augmented prompt sent to the delegated agent |
launch.sh |
The generated executable launcher script |
done.md |
The completion summary written by the delegated agent |
In tmux mode, agent-offload appends instructions to the prompt telling the
delegated agent to write a concise summary to done.md.tmp, then atomically
rename it to done.md. The parent process waits for done.md to exist. If the
tmux pane closes first, the run fails instead of hanging silently.
The delegated pane opens on the right side of the tmux window that runs
agent-offload, even if another tmux client is viewing a different window.
Configuration
A config has a default profile and one or more named profiles. This example uses Codex Spark and a DeepSeek-backed Claude Code profile:
default_profile: claude-deepseek-flash
profiles:
codex-spark:
command: /opt/homebrew/bin/codex
interface: codex
args:
- --yolo
- --model
- gpt-5.3-codex-spark
- -c
- model_reasoning_effort="high"
env:
prompt: argument
claude-deepseek-flash:
command: /Users/raine/.local/bin/claude
interface: claude
args:
- --dangerously-skip-permissions
env:
ANTHROPIC_BASE_URL: https://api.deepseek.com/anthropic
ANTHROPIC_AUTH_TOKEN:
from_env: DEEPSEEK_API_KEY
ANTHROPIC_MODEL: 'deepseek-v4-flash[1m]'
ANTHROPIC_SMALL_FAST_MODEL: 'deepseek-v4-flash[1m]'
ANTHROPIC_DEFAULT_OPUS_MODEL: deepseek-v4-flash
ANTHROPIC_DEFAULT_SONNET_MODEL: deepseek-v4-flash
ANTHROPIC_DEFAULT_HAIKU_MODEL: deepseek-v4-flash
CLAUDE_CODE_SUBAGENT_MODEL: deepseek-v4-pro
CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC: '1'
CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK: '1'
CLAUDE_CODE_EFFORT_LEVEL: max
CLAUDE_CODE_ENABLE_TELEMETRY: '0'
prompt: argument
cursor-composer:
command: /Users/raine/.local/bin/cursor-agent
interface: cursor
args:
- --force
- --model
- composer-2.5-fast
env:
Profile fields
| Field | Description |
|---|---|
command |
Binary or script to execute |
interface |
Argument shape for the agent CLI |
args |
Extra arguments passed before the prompt |
env |
Environment variables exported before launch |
prompt |
How the augmented prompt is delivered |
headless |
Whether to run this profile without tmux |
Interfaces
| Interface | Prompt argument form |
|---|---|
generic |
-- "$PROMPT_CONTENT" |
claude |
-- "$PROMPT_CONTENT" |
codex |
-- "$PROMPT_CONTENT" |
cursor |
-- "$PROMPT_CONTENT" |
opencode |
--prompt "$PROMPT_CONTENT" |
Prompt delivery
| Value | Behavior |
|---|---|
argument |
Pass the prompt using the selected interface's argument form |
stdin |
Redirect the prompt file to stdin |
prompt-file-arg |
Replace {prompt_file} in args with the prompt file path |
Use prompt-file-arg only when at least one arg contains {prompt_file}.
Environment variables
Literal values are written directly:
env:
CLAUDE_CODE_ENABLE_TELEMETRY: '0'
Values can also be forwarded from the host environment:
env:
API_KEY:
from_env: MY_SECRET_KEY
Environment variable names must be valid shell identifiers.
Commands
run
Launch a profile with a prompt and wait for completion.
|
The run subcommand is optional, so this is equivalent:
Options:
| Option | Description |
|---|---|
-p, --profile <name> |
Profile name from the config |
--config <path> |
Override the config file path |
-H, --headless |
Run this invocation in headless mode |
profiles
List configured profiles and mark the default.
install-skill
Install the bundled skill for one provider or all detected providers.
The command is idempotent. It prints up-to-date when the installed skill
matches the bundled copy.
Agent workflow
A typical host-agent workflow looks like this:
> /agent-offload implement the plan in history/2026-06-06-plan-feature.md with the spark profile
The host agent:
1. Loads the `agent-offload` skill
2. Pipes the plan into `agent-offload run --profile codex-spark`
3. Waits for the delegated agent's completion summary
4. Inspects the diff and runs the requested checks
5. Reports the reviewed result
The delegated agent should do the implementation. The host agent should still inspect the diff, verify behavior, and decide what to report or commit.
Requirements
- Rust, for building from source
- tmux
- At least one configured agent CLI
Development
just check runs formatting, clippy fixes, build, tests, and clippy.
Useful commands:
Related projects
- consult-llm - ask other LLMs for planning, review, and debugging help
- git-surgeon - non-interactive hunk staging and commit surgery for agents
- workmux - git worktrees and tmux windows for parallel agent workflows