Multi-Step Orchestration

Recipes & Workflows

Declarative multi-step agent coordination with approval gates, loops, and automatic checkpointing

Active
February 2026
The Problem

Why recipes exist

Manual Orchestration
Complex workflows involve multiple agents in sequence. Coordinating them by hand is error-prone and tedious.
No Repeatability
You run the same multi-step workflow repeatedly. Without a definition, each run varies. Encoding it as YAML makes it consistent and versioned.
Lost Progress
Long-running work gets interrupted. Without checkpointing, you start over. Recipes resume exactly where they stopped.
Anatomy

Declarative YAML workflows

name: code-review-workflow description: Review code and deploy on approval steps: - id: analyze agent: foundation:explorer output: analysis prompt: "Analyze changes in {{file_path}}" - id: review agent: foundation:zen-architect output: review_result prompt: "Review code quality: {{analysis}}" - id: approve type: approval_gate message: "Review complete. Deploy?" - id: deploy type: bash command: "deploy.sh {{file_path}}"
Capabilities

Core features

Context Accumulation
Each step's output field becomes a named variable. Reference it in later steps with {{variable}} syntax. Context flows forward automatically.
Approval Gates
Pause execution for human review at critical checkpoints. The workflow waits until you approve or deny before proceeding.
Automatic Checkpointing
Every step completion is persisted. If a workflow is interrupted, resume it later with all previous results intact.
Advanced

Powerful patterns

Foreach + Parallel
Iterate over collections with foreach. Add parallel: N to process N items concurrently. Ideal for multi-file analysis.
While / Convergence
Loop with while_condition until a quality threshold is met. Set max_iterations as a safety bound.
Conditional Execution
Skip steps based on conditions. Route workflows dynamically: skip deep validation for clean code, escalate critical findings.
Human-in-Loop

Approval checkpoints

Agent: Analyze
Agent: Review
⏸ Approval Gate
↓ human approves
Bash: Deploy
Critical checkpoints. Review before deployment, approve before destructive operations, inspect results before committing. Staged recipes group steps into phases with approval between each stage.
Fault Tolerance

Resume where you left off

# Recipe interrupted at step 3? Resume it: amplifier run "resume recipe session recipe_20260121_abc123" # Or list all active sessions: amplifier run "list recipe sessions" # All previous step results preserved # Context accumulation intact # Picks up at the exact step that was interrupted
Long-running workflows. Close your laptop, go to a meeting, come back and resume. The recipe engine checkpoints after every step completion.
Velocity

Built in 44 days

74
Commits
10,156
Lines of Python
46
Example recipes
Dec 15, 2025 – Jan 28, 2026 · 5 contributors · Brian Krabach: 68 of 74 commits (~92%)
Sources

Research Methodology

Data as of: February 20, 2026

Feature status: Active

Repository: microsoft/amplifier-bundle-recipes

Research performed:

Primary contributor: Brian Krabach (68 of 74 commits, ~92%)

Other contributors: Marc Goodner (3), Mollie Munoz (1), Sam Schillace (1), momuno (1)

Gaps: No performance benchmarks available. No usage telemetry. Impact claims are qualitative.

Automate your workflows

Declarative, resumable, repeatable

Create recipes: Use the recipe-author agent — don't write YAML directly

Examples: 46 templates in amplifier-bundle-recipes/examples/

Run a recipe: amplifier run "execute @recipes:examples/code-review-recipe.yaml"
More Amplifier Stories