Generate complete Agent Workflow Protocol (AWP) compliant multi-agent workflows from natural language descriptions. Produces workflow.awp.yaml, agent configs, prompts, schemas, and optionally custom tools and skills.
AWP organizes a workflow into seven protocol layers. Each layer answers one question:
<svg viewBox="0 0 620 290" xmlns="http://www.w3.org/2000/svg" font-family="sans-serif" font-size="13"> <rect x="20" y="5" width="580" height="34" rx="4" fill="#dce6f7" stroke="#4a6fa5" stroke-width="1.5"/> <text x="30" y="27" font-weight="bold" fill="#2a3f5f">Layer 6</text> <text x="120" y="27" fill="#2a3f5f">OBSERVABILITY</text> <text x="590" y="27" text-anchor="end" fill="#666">How do I monitor this workflow?</text> <rect x="20" y="43" width="580" height="34" rx="4" fill="#d5e8d4" stroke="#5b8c5a" stroke-width="1.5"/> <text x="30" y="65" font-weight="bold" fill="#2d5a2d">Layer 5</text> <text x="120" y="65" fill="#2d5a2d">ORCHESTRATION</text> <text x="590" y="65" text-anchor="end" fill="#666">In what order and under what conditions?</text> <rect x="20" y="81" width="580" height="34" rx="4" fill="#fef3cd" stroke="#d4a017" stroke-width="1.5"/> <text x="30" y="103" font-weight="bold" fill="#856404">Layer 4</text> <text x="120" y="103" fill="#856404">MEMORY & STATE</text> <text x="590" y="103" text-anchor="end" fill="#666">What does the workflow remember?</text> <rect x="20" y="119" width="580" height="34" rx="4" fill="#e8d5f5" stroke="#7b4ea3" stroke-width="1.5"/> <text x="30" y="141" font-weight="bold" fill="#5a2d82">Layer 3</text> <text x="120" y="141" fill="#5a2d82">COMMUNICATION</text> <text x="590" y="141" text-anchor="end" fill="#666">How do agents talk to each other?</text> <rect x="20" y="157" width="580" height="34" rx="4" fill="#fde2e2" stroke="#c0392b" stroke-width="1.5"/> <text x="30" y="179" font-weight="bold" fill="#922b21">Layer 2</text> <text x="120" y="179" fill="#922b21">CAPABILITIES</text> <text x="590" y="179" text-anchor="end" fill="#666">What can an agent do? (tools, skills)</text> <rect x="20" y="195" width="580" height="34" rx="4" fill="#d5f5e3" stroke="#27ae60" stroke-width="1.5"/> <text x="30" y="217" font-weight="bold" fill="#1a6b3c">Layer 1</text> <text x="120" y="217" fill="#1a6b3c">AGENT IDENTITY</text> <text x="590" y="217" text-anchor="end" fill="#666">Who is this agent?</text> <rect x="20" y="233" width="580" height="34" rx="4" fill="#f0f0f0" stroke="#888" stroke-width="1.5"/> <text x="30" y="255" font-weight="bold" fill="#333">Layer 0</text> <text x="120" y="255" fill="#333">MANIFEST</text> <text x="590" y="255" text-anchor="end" fill="#666">What is this workflow?</text> </svg>You start at the bottom. Layer 0 (manifest) and Layer 1 (agent identity) are always required. Everything above is opt-in.
Autonomy levels measure HOW AUTONOMOUS the workflow is, not WHAT FEATURES it has. Communication, memory, and observability are cross-cutting features available at any level.
| Level | Name | Description |
|---|---|---|
| A0 | Prescribed | Static DAG, predefined agents, fixed tools. Minimum viable workflow. |
| A1 | Adaptive | Conditional execution, loops, fan-out, multi-agent DAG. |
| A2 | Delegating | Manager spawns workers dynamically (delegation loop). Budget required. |
| A3 | Self-Tooling | Agents create tools and skills at runtime. Safety envelope required. |
| A4 | Self-Organizing | Recursive delegation, budget distribution. Observability required. |
Cross-cutting features (any level): Communication, Memory, Observability, Security.
| Feature | Config Location | Description |
|---|---|---|
| Agent DAG | workflow.awp.yaml graph | Directed acyclic graph of agents with depends_on edges. |
| State Sharing | graph[].share_output | Fields from an agent's output available to dependents. |
| Execution Modes | execution.mode | sequential, parallel, or conditional execution. |
| MCP Tools | agent.awp.yaml tools | Tool calling via the MCP registry (web, file, shell, etc.). |
| Message Bus | communication.bus | In-memory message passing between agents. |
| Memory | memory.tiers | Long-term (MEMORY.md) and working (daily logs) memory. |
| Skills | workflow skills/ | Markdown knowledge injected into agent system prompts. |
| Preprocessor | agent workflow/preprocessor/ | Data extraction and feature engineering before LLM call. |
| Vision | agent.awp.yaml vision | Image processing via base64-encoded data URLs. |
| Archetypes & Recipes | runtime, no YAML | Six generic tool archetypes (compute / fetch / parse / transform / render / probe). The manager plans capabilities as reuse_or_generate: "synthesize" with archetype_id + recipe_params, the runtime synthesises a deterministic handler, and successful instantiations are auto-captured as content-addressed Recipes under ~/.awp/recipes/ (Quarantined → Probationary → Trusted) for replay-gated reuse on future runs. Replaces the old hand-rolled pattern library as the primary path for new tools — see R31 below. |
AWP provides two orchestration engines. The choice determines how agents coordinate:
DAG Engine (default, A0-A1): Static graph of agents. Dependencies defined in YAML. Topological execution.