Skip to main content

​
Overview

With Subagents, the master agent spawns isolated sub-processes at runtime. Each Subagent has its own fresh context, model, and skill — it doesn’t carry the conversation history of the master or previous Subagents. This enables parallel execution (e.g., three reviewers running simultaneously) and keeps the master agent’s context clean throughout the pipeline.
Presets vs. Subagents — which to use: Use Agent Presets when steps run sequentially and you want model assignments configured once in Settings and reused across tasks. Use Subagents when you need parallel execution within a phase, clean context isolation between phases, or want model assignments to live in the workflow file itself.
See the Subagents page for the full reference on how Subagents work.

​
The Three-Phase Pipeline

auto and auto-plus are Zenflow’s managed model selectors. auto routes to a fast, cost-efficient model; auto-plus routes to a frontier reasoning model. Both are updated automatically as better models become available.

​
Why This Structure Works

Phase 1 — Planner (frontier reasoning model): Architectural decisions made here flow into every downstream phase. Investing in a high-quality plan pays dividends: the implementer doesn’t need to re-derive architecture decisions, and the reviewer has a concrete reference to evaluate against. Phase 2 — Implementer (fast model): Implementation is the most token-intensive phase. A fast model executing a well-defined spec produces equivalent quality to an expensive model working without one — because the spec eliminates the need for architectural exploration. This is where token count is highest and where model selection has the most impact. Phase 3 — Review Orchestrator (parallel, diverse models): The orchestrator spawns three workers simultaneously, each reviewing the diff from a different model’s perspective. Cross-model review catches a broader class of issues than same-model review. The orchestrator synthesizes findings into a single consolidated report.

​
Built-In Skills

Zenflow ships built-in skills for each phase: You specify the model; the skill handles the task structure and output format. You can also trigger the parallel review pattern from chat at any point, without a full pipeline:
This uses Subagents to run multiple models against the same diff in parallel and synthesize their findings into a single review.