Multi Agent Pipeline

Multi-agent pipeline

🌍 English translation is auto-generated on release. See the French version for the current complete page.

The pipeline chains specialized agents: Spec Creator → Planner → Coder → QA Reviewer → QA Fixer. This page details each phase, the data they exchange, and tier-based phase counts.

Multi-Agent Pipeline

At the core of WorkPilot AI is an autonomous pipeline that carries a task from description to reviewed pull request without manual hand-off between stages. Once a user creates a task, a spec creation pipeline assesses complexity and writes a specification, which then flows through four specialized agents:

Stage	Agent	Responsibility
1	Planner	Complexity assessment, phased subtask breakdown, dependency analysis
2	Coder	Context-aware implementation, capable of spawning parallel subagents
3	QA Reviewer	Validates the implementation against acceptance criteria
4	QA Fixer	Automatically resolves issues the reviewer identifies

The QA stage runs as an auto-fix loop: the Reviewer and Fixer cycle until acceptance criteria pass (up to 50 iterations via the CLI), so failing checks are addressed without requiring a human to intervene at every round. After QA approval, the work moves to human review before merging.

Every task executes in an isolated git worktree, keeping the main branch untouched until the user explicitly reviews and merges — and an AI-powered semantic merge step handles conflict resolution when integrating changes back. The pipeline itself is now describable declaratively: a workflow.yaml file lays out build phases, and an engine resolves which phases actually run based on the chosen effort level, the AI provider's capabilities, and the files a task touches — with a "Resolved Profile" preview showing users what their effort setting buys before a build starts.

For larger workloads, the pipeline scales horizontally too: up to 12 agent terminals can run in parallel, and Mission Control provides a live, NASA-style dashboard for monitoring multiple agents' status, token usage, and reasoning simultaneously. Individual phases (planning, coding, validation) can also be re-run on demand without restarting the entire task, and a running build can be paused and resumed later — optionally on a different LLM provider — with full context preserved.

---

➡️ Continue: Specialized agents