Ultra-lightweightMulti-agent OrchestratorAnywhere you work

Write plan. Run plan. Workbench does the rest.

$brew install duncankmckinnon/tap/workbench

Get Started View on GitHub Download Skills

terminal

$ wb run my-feature

Parsed 4 task(s) from .workbench/my-feature/plan.md

Plan: My Feature

Tasks: 4 across 2 wave(s)

Concurrency: 4

Max retries: 2

Session branch: my-feature

tmux: enabled

Workbench

Task	Wave	Status	Branch	Time	Pipeline	Merged
User model	1	done	wb/user-model	3m42s	impl:ok → test:pass → review:pass	✓
Auth utils	1	done	wb/auth-utils	2m05s	impl:ok → test:pass → review:pass	✓
API routes	2	done	wb/api-routes	4m18s	impl:ok → test:fail → fix → test:pass → review:pass	✓
Web UI	2	done	wb/web-ui	5m10s	impl:ok → test:pass → review:pass	✓

━━━ Summary ━━━

✓ 4 completed

All changes merged into: my-feature

git checkout my-feature

git diff main...my-feature

How it works

Write a plan

Describe tasks in a markdown file with dependencies and file ownership

Run one command

wb run plan.md dispatches agents in parallel

Ship the result

Agents implement, test, review, and merge — you review the PR

The agent pipeline for each task

test

implement

test

fix

review

fix

merge

The agent pipeline on completion

collect requirements

comprehensive review

fix

merge

create PR

From plan to code in minutes

.workbench/myfeature/plan.md

---

name: myfeature

max_concurrent: 4

tdd: true

final_review: true

---

## Context

Building a REST API with Express and TypeScript...

## Task: User model

Files: src/models/user.ts, src/db/schema.ts

Create the User model with email, password hash,

and session fields. Add migration.

## Task: Auth endpoints

Files: src/routes/auth.ts

Depends: user-model

Implement login, register, and logout

endpoints using the User model.

terminal

$ wb run myfeature --final-review

Parsed 2 task(s) from .workbench/myfeature/plan.md

1. User model (src/models/user.ts, src/db/schema.ts)

2. Auth endpoints (src/routes/auth.ts)

Plan: My Feature

Tasks: 2 across 2 wave(s)

Concurrency: 4

Max retries: 2

Repo: ~/code/myfeature

Session branch: myfeature

tmux: enabled

Workbench

Task	Wave	Status	Branch	Time	Pipeline	Merged
User model	1	done	wb/user-model	2m05s	test:fail → impl:ok → test:pass → review:pass	✓
Auth endpoints	2	done	wb/auth-endpoints	4m18s	test:fail → impl:ok → test:pass → review:fail → fix → review:pass	✓

━━━ Summary ━━━

✓ 2 completed

All changes merged into: myfeature

git checkout myfeature

git diff main...myfeature

Post-task agents

Agent	Status	Time	Output / Note
Summarizer	done	1m40s	.workbench/myfeature/wrap-up/myfeature/requirements.md
Branch reviewer	done	3m20s	.workbench/myfeature/wrap-up/myfeature/review.md
PR writer	done	0m52s	https://github.com/you/myfeature/pull/42

Everything you need to orchestrate AI agents

Agent Skills

Bundled skills let your agent develop plans and guide the whole process — install workbench, write plans, configure agents, and generate conventions

Multi-Agent Support

Assign any supported agent CLI — Claude Code, Gemini, Codex, Cursor, Copilot, or custom adapters — to each pipeline role

Automated Review & PRs

Requirements digest up front, branch review at the end, and a polished pull request description on completion

Parallel Execution

Independent tasks run simultaneously in isolated git worktrees, ordered into waves by their dependencies

Granular Step Control

Run a single wave or task, skip tests or review, and control merge and PR steps independently

TDD Mode

Tests first, implementation second — built into the pipeline

Plans by Name

Reference plans by name during runs (wb run <plan-name>) — frontmatter configures the run, and every artifact lives in .workbench/<plan-name>/

Shared Conventions

Add or generate a conventions.md in your .workbench folder so every plan inherits the same instructions for style, formatting, and more

Auto-Merge

Parallel branches merged automatically with conflict resolution