deepwork

Orchestrates phased multi-agent coding workflows with Oracle review gates and progress tracking.

1|Updated Jul 23, 2026
One-click install
npx skills add https://github.com/sanjanb/my-agent-harness --skill deepwork-sanjanb
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: deepwork
Source: https://github.com/sanjanb/my-agent-harness/tree/main/skills/deepwork
Command: npx skills add https://github.com/sanjanb/my-agent-harness --skill deepwork-sanjanb

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Large, high-risk coding efforts with multiple dependent phases often fail due to poor coordination, lost context, and unreviewed intermediate work. This Skill structures those efforts as scheduled, phase-gated orchestration with persistent progress state. ## Core Features & Use Cases - Phased Delegation: Splits work into coherent implementation phases, assigns bounded scopes to specialist agents, and keeps OpenCode todos aligned with the active phase. - Oracle Review Gates: Runs an Oracle review after each planned phase, batches actionable findings into one remediation pass, and only re-reviews when the decision or risk changes. - Persistent Progress State: Maintains a markdown file under .slim/deepwork/ capturing research findings, plans, review notes, and validation results across the session. - Parallel Worktree Execution: Uses worktree isolation and a task board for parallel agents, with strictly sequential verified merges. - Use Case: When migrating a cross-cutting architecture that touches many modules, activate deepwork to plan phases, delegate to specialists like @designer and @fixer, gate each phase with Oracle review, and track everything in a local progress file. ## Quick Start Use the deepwork skill to plan and orchestrate this large multi-phase refactoring with specialist delegation and Oracle review gates.

Frequently Asked Questions about deepwork

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I orchestrate a large multi-phase coding task with AI agents?

Activate deepwork to draft a plan, split the work into coherent phases based on dependencies, delegate bounded scopes to specialist agents, and gate each phase with an Oracle review. Progress is tracked in a markdown file under .slim/deepwork/.

When should I use deepwork instead of a normal coding workflow?

Use deepwork only for large or high-risk work: multiple dependent phases, cross-cutting architectural change, or unsafe-to-partially-ship migrations. Do not use it for trivial edits, quick docs changes, simple bug fixes, or routine bounded features.

How do Oracle review gates work between implementation phases?

After each planned phase, the orchestrator validates results, updates the deepwork file, and asks Oracle to review before continuing. Actionable findings are batched into one remediation pass, and a follow-up review happens only if the decision or risk changed.

Can multiple agents work in parallel during a deepwork phase?

Yes. Independent agents each get an isolated git worktree and a registered task, then their branches are merged sequentially into main with verification after each merge. Simultaneous merges are never allowed.

What are the limitations of the deepwork orchestration approach?

Deepwork adds planning, review, and state-tracking overhead, so it is wasteful for routine multi-file changes or small tasks. It also depends on specialist agents like @oracle, @designer, and @librarian being available in the environment.