Open source orchestration layer

Not another agent.
The quality layer
on top of them.

DeepWork adds quality gates, self-critique loops, and learning to Claude Code, Aider, and Cursor. Your agents do the work. DeepWork makes sure it's finished work.

Open source CLI-native 10 min to first workflow
$ brew install unsupervisedcom/deepwork/deepwork
$ deepwork define api-migration
Created job.yml + 4 skill files
$ deepwork run api-migration
── step 1/4: extract endpoints ──
Mapped 23 endpoints across 6 route files
── step 2/4: generate new schema ──
Self-critique: missing pagination params. Re-running.
── step 3/4: write migration code ──
Generated 1,847 lines with backward compat layer
── step 4/4: quality gate ──
Score: 91/100 ✓ Passed. Output saved.
$ deepwork learn "always include rate limit headers"
Skill updated. Future runs will include rate limiting.

AI coding agents are powerful.
But power ≠ production-ready.

Three frustrations every developer hits once they rely on AI agents for real work.

📉

Quality degrades in long sessions

The first output is great. By step 20, the agent is hallucinating variable names and forgetting constraints. There's no checkpoint. No validation. Just drift.

70% of developers report quality decay in extended sessions
🎲

Inconsistent output across same tasks

Run the same prompt twice, get two different results. No way to enforce structure, style, or correctness across repeated workflows. You review everything, every time.

Same task, different output — no repeatability
🔗

No way to chain multi-step workflows

Complex work needs sequential stages with validation between each step. Current agents treat everything as a single-pass operation. Break it into steps manually, or hope for the best.

Developers spend 30% of time wiring context, not coding

Define → Execute → Learn

Three primitives. Infinite workflows. DeepWork adds structure to autonomy without taking away any power.

01 / Define

Multi-step jobs with quality gates

Break complex work into stages. Each step gets its own SKILL.md file — YAML frontmatter for config, markdown for instructions. The agent executes sequentially, validating at each checkpoint before moving forward.

02 / Execute

Self-critique + retry loops

Each step includes a critique pass where the agent reviews its own output, scores it 0–100, catches errors, and re-runs until the quality bar is met. Fire and forget — check back for validated output.

03 / Learn

Automated self-improvement

Run deepwork learn with feedback and skills update automatically. Every future run incorporates what you taught it. Quality compounds over time.

The loop, live

Competitive research job: quality gates running, a retry when the bar isn't hit, and the learn cycle at the end.

deepwork — competitive-research

Same agent. Better results.

Claude Code is the engine. DeepWork is the quality layer. You don't replace your agent — you make it finish what it starts.

Raw Claude Code

  • Single-pass output, hope it's good
  • Manual review every time
  • Context lost between sessions
  • Same mistakes on repeat tasks
  • No way to enforce quality bar
  • You steer, the agent types

Claude Code + DeepWork

  • Quality gates validate each step
  • Self-critique catches errors automatically
  • Skills persist context across all runs
  • Learn function eliminates repeat mistakes
  • Scores 0–100, retries until passed
  • You set direction, the agent delivers
Quality Gates

Per-step validation, not post-hoc review

Every stage gets scored. Failures trigger automatic retries with the critique feedback injected. Output quality is enforced, not hoped for. Read how quality gates work →

Self-Critique Loops

The agent reviews its own work

Before any step passes, the agent critiques its output against the skill spec. Catches hallucinations, missing requirements, and structural issues before you ever see them.

Background Execution

Fire and forget, check back later

Define the job, run it, go do something else. DeepWork handles the multi-step execution, validation, and retry logic. You get the finished output. Build your first workflow →

Learning from Feedback

Skills improve every time you use them

One command updates the skill file with your feedback. The next run is better than the last. Quality compounds — your workflows get smarter over time.

Open by default

Not another closed platform with surprise rate limits. DeepWork is built for developers who want control.

💰

Transparent pricing

DeepWork is free and open source. You pay only for the model tokens you use — at the provider's price, not ours. No subscriptions, no hidden quotas.

🔀

Multi-model flexibility

Use Claude, GPT, Gemini, or any model per task. Pick the right model for the right step. Cheap reasoning for boilerplate, top-tier for architecture decisions.

👥

Community-first governance

Open source under BSL 1.1. Transparent roadmap. No VC collapse risk. Skills are composable and shareable — build once, use everywhere.

Developers are shipping with DeepWork

Reported productivity gains
from early users
10 min
From install to first
validated workflow
83%
Baseline quality gate
pass rate, improving with learn

Install DeepWork

Two commands. Ten minutes to your first validated workflow.

brew tap unsupervisedcom/deepwork && brew install deepwork click to copy