Skip to main content

The Harness Engineering Pipeline

Every task in Maestro flows through a deterministic pipeline. Each stage is handled by a specialized agent, and the pipeline enforces quality gates between stages.

Pipeline stages​

Queued​

A task enters the queue when it's created — either manually from the dashboard, via the API, or synced from a GitHub issue or Linear ticket.

What happens:

  • Task is validated and assigned to a workspace
  • The task receives a unique internal ID
  • Status is set to queued

Implement​

The implementation agent picks up the task and writes code.

What happens:

  • Agent reads the project codebase to understand context
  • Writes the implementation based on the task description
  • Creates a Git branch and opens a pull request
  • Attaches the PR link to the task

Review​

The review agent reads the pull request diff and performs an inline code review.

What happens:

  • Agent reads the full diff of the PR
  • Leaves inline comments on specific lines where it finds issues
  • Comments cover: bugs, missing validation, style, performance, security
  • If no issues are found, the review is approved immediately

Review loop​

If the review agent leaves comments, the task cycles back to the implementation agent:

  1. Implementation agent reads the review comments
  2. Applies fixes and pushes new commits
  3. Task returns to the review agent for re-review
  4. This loop continues until all comments are resolved

The loop has a configurable maximum iteration count (default: 5) to prevent infinite cycles.

Risk Profile​

After the review is approved, the risk profile agent analyzes the change.

What happens:

  • Scores the change across multiple dimensions:
    • Complexity — how many files, lines, and logical branches are affected
    • Blast radius — how many other modules depend on the changed code
    • Test coverage — whether tests exist for the new/changed code
  • Produces an overall risk score: LOW, MEDIUM, or HIGH
  • LOW tasks can be auto-merged; HIGH tasks require human approval

Deploy​

The deployment agent handles merging and CI.

What happens:

  • Merges the PR into the target branch
  • Monitors CI pipeline status
  • Reports success or failure back to the task

Monitor​

Post-deployment monitoring (when configured).

What happens:

  • Watches application logs and metrics for anomalies
  • If errors spike after deployment, the task is flagged for rollback review
  • Reports monitoring status back to the task

Status transitions​

queued → implement → review ↔ implement (loop) → risk_profile → deploy → monitor → done

At any point, a task can transition to failed if an unrecoverable error occurs, or blocked if human intervention is required.