How It Works

The Pipeline

Objective
    │
    ▼
 Planner ──► Plan (streams, file scopes, dependencies)
    │
    ▼
 Human Approval
    │
    ▼
 Dispatch ──► Parallel Agents in Isolated Worktrees
    │
    ├── Stream 1: Build ──► Quality Gates ──► Merge Ready
    ├── Stream 2: Build ──► Quality Gates ──► Merge Ready
    └── Stream 3: Build ──► Quality Gates ──► Merge Ready
    │
    ▼
 Merge Queue ──► Post-Merge Gates ──► PR Created
    │
    ▼
 Objective Complete

Each step is either deterministic (quality gates, merge, scheduling) or agentic (planning, building, reviewing). The blueprint YAML defines which is which. You control the workflow.

Agent Roles

Role	What it does
Planner	Explores the codebase, decomposes the objective into parallel streams with file scopes and dependencies
Builder	Implements the stream's task in an isolated worktree, commits changes
Reviewer	Reviews the implementation for correctness and quality

Key Principles

Deterministic infrastructure. Everything that isn't an LLM decision is deterministic: scheduling, gates, merge, scoped rules.
Isolation by default. Each agent works in its own sandbox. Agents never see each other's changes during execution.
Human in the loop. You approve the plan before execution begins. Escalations pause for your input.
Partial completion. When some streams fail, Tack merges what succeeded. You can retry failures without re-running everything.

How It Works

The Pipeline

Agent Roles

Key Principles

On this page