awman

---

awman (Agent Workflow Manager) is a developer tool that adds structure and automation to the whole agentic software development lifecycle: from issue to merged PR.

4 stages of improved agentic software development with awman

1. Isolate your code agents with containers and worktrees 🛑
2. Run multiple agents in parallel with the TUI 🔄
3. Turn your team's development lifecycle into repeatable workflows 📈
4. Automate the rest — build a squad of agents to tackle recurring work, fan out to your homelab or cluster with API mode 🤝

---

Installation

curl -s https://prettysmart.dev/install/awman.sh | sh

The installer detects your platform and puts awman on your PATH.

With mise — using the GitHub backend:

mise use -g github:prettysmartdev/awman

To pin to a specific version: mise use -g github:prettysmartdev/[email protected]

From GitHub Releases — download the binary for your platform from GitHub Releases:

Platform	Asset
Linux (x86_64)	awman-linux-amd64
Linux (ARM64)	awman-linux-arm64
macOS (Apple Silicon)	awman-macos-arm64
Windows (x86_64)	awman-windows-amd64.exe

The Apple Silicon asset includes the experimental builtin microVM runtime and is ad-hoc signed (not notarized). If you download it with a browser, macOS quarantines it: verify its SHA-256 against the release page, then run xattr -d com.apple.quarantine awman-macos-arm64. The installer and mise do not quarantine it. See Runtimes: The macOS binary, Gatekeeper and permissions.

From source — requires Rust 1.94+ and make:

git clone https://github.com/prettysmartdev/awman.git
cd awman
sudo make install

---

Quick Start

# 1. Initialize your repo (once per project)
awman init

# 2. Open the TUI
awman

# 3. Start an agent session
chat

# 4. Optionally run the Dockerfile.dev refresh agent to
#    ensure all your project's tools get installed
ready --refresh

See the Getting Started Guide for a full walkthrough.

---

What you can do with awman

Run multiple agents at once

Open new tabs in the TUI with Ctrl+T. Each tab is independent — its own working directory, its own container — and keeps running in the background while you work in another tab. Switch with Ctrl+A / Ctrl+D.

When an agent goes quiet for 30 seconds, whether it's stuck or just finished and waiting on you, its tab turns yellow so you know where to look.

Run structured workflows

A workflow breaks complex work into phases — plan → implement → review → docs. Each phase is a separate agent session in its own container, so you review the output between phases and decide whether to continue, retry, or redirect.

Workflows are TOML or YAML files in your repo. Steps declare their dependencies with depends_on, and independent steps run in parallel. Around them, setup and teardown phases prepare the branch and handle the finish: run tests, commit, push, open the PR, and block on CI until it goes green. Any of those can carry an on_failure block that launches an agent to fix the problem and retries.

Three things make this more than a prompt runner:

• Per-step agents. Any step can name the agent and model it runs with, so you can have Codex implement and Claude review in the same pipeline. Steps that don't specify one use your project's default.
• Per-step overlays. Each step declares exactly what it needs from the host — skill(review) for a review checklist, env(API_KEY) for a tool it calls. Nothing gets host access it didn't ask for.
• Prerequisite checks. Before anything starts, awman checks that commit, push, PR and CI steps have the Git identity, SSH key or GitHub token they need, and asks before handing a credential to the step that uses it.

awman exec workflow ./aspec/workflows/implement-pr.toml --work-item 0027

The --work-item is optional: pass one to substitute a spec you've written into the prompts, or leave it off. See Workflows for the full file format, template variables, and the control board.

When a step's agent dies, the workflow doesn't. The control board opens with the failure attached so you can retry the step, step back to the one before it, step over it, or pause — and a paused run picks up from a named step the next time you start it. Unattended runs (squad, the API server) retry a failed step once on their own before giving up.

title = "Implement Feature"

[[setup]]
type = "checkout_create_branch"
branch = "feature/{{work_item_number}}"
base = "main"

[[step]]
name = "plan"
prompt = "Read work item {{work_item_content}} and produce an implementation plan."

[[step]]
name = "implement"
depends_on = ["plan"]
agent = "codex"
prompt = "Implement work item {{work_item_number}} according to the plan."

[[step]]
name = "review"
depends_on = ["implement"]
agent = "claude"
prompt = "Review the implementation for correctness and style."
overlays = ["skill(review)"]

[[teardown]]
type = "run_shell"
command = "make test"

[[teardown]]
type = "commit_changes"
message = "Implement {{work_item_number}}"

[[teardown]]
type = "push_branch"

[[teardown]]
type = "create_pull_request"
title = "Implement {{work_item_number}}"

[[teardown]]
type = "poll_ci"

Supported agents: claude, codex, opencode, gemini, antigravity, copilot, cline, plus any custom agent you define (see Custom Agents; crush and maki ship as examples).

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗