A Practical A2A Protocol for Supply Chain Intelligence

A lightweight pattern for exposing modular capabilities without creating a monolith

Note: In this article, “agents” refers to modular capability units — not LLM orchestrators or swarm-style multi-agent systems.

Supply chain operations rarely fit inside a simple request–response cycle.

Classification, tariff evaluation, dependency mapping, due diligence checks — all of them unfold across multiple dependent steps. Yet most API designs still assume these workflows are instantaneous.

While developing large-scale supply chain intelligence systems—multi-tier supplier graphs, risk propagation logic, compliance engines, tariff models—a recurring challenge emerged:

How do you expose complex, multi-step logic without forcing complexity onto the caller?

A lightweight A2A (Agent-to-Agent) protocol proved to be an effective solution — not as a product feature, but as a generalizable integration pattern for long-running or multi-step API workflows.

This article breaks down the reasoning behind the pattern and why it works well in practice.

Why Traditional APIs Struggle With Multi-Step Workflows

On the surface, workflows like “classify a product” or “calculate tariff duties” look atomic. In reality, they decompose into multiple phases:

• input normalization
• attribute inference
• rule and exception matching
• multi-country tariff overlays
• multi-hop dependency expansion
• evidence collection and assembly

Bundling these processes into a single synchronous call leads to:

1. unpredictable execution times
2. brittle timeout handling
3. opaque branching logic
4. unclear error reporting
5. monolithic endpoints that quickly become unmanageable

The issue isn’t performance.

It’s the lack of an API abstraction that expresses multi-step workflows cleanly.

Why a Lightweight A2A Pattern Works

Instead of a massive all-in-one endpoint, a system can be decomposed into small, focused capability units, each responsible for exactly one task:

• classification
• tariff calculation
• dependency graph expansion
• concentration analysis
• due diligence synthesis

Each capability exposes a simple lifecycle:

1. run — start the task
2. status — check progress
3. result — fetch the final structured output

This keeps the surface area predictable while allowing internal logic to remain expressive.

The A2A Lifecycle (with Examples)

A2A follows three predictable operations:

caller → POST /run      → task_id
caller → GET /status    → pending | running | done | failed
caller → GET /result    → structured output

This minimal protocol is intuitive, flexible, and maps naturally to complex workflows.

1. run — Start a task (idempotent)

POST /agents/hts_classification/run
{
  "product_description": "EV lithium battery pack",
  "idempotency_key": "client-123"
}

Response:

{
  "task_id": "abc123",
  "status": "PENDING"
}

Idempotency keys allow safe retries — essential in distributed environments.

2. status — Monitor progress

GET /agents/hts_classification/status?task_id=abc123

Response:

{
  "status": "RUNNING",
  "progress": 0.4,
  "steps": ["normalize_input", "attribute_inference"]
}

For multi-second or multi-minute workflows, clients typically poll every 1–3 seconds.

3. result — Retrieve structured output

GET /agents/hts_classification/result?task_id=abc123

Response:

{
  "hts_code": "8507.60.0020",
  "confidence": 0.94,
  "evidence": {
    "matched_phrases": ["lithium-ion", "battery pack"],
    "rule_path": ["chapter 85", "heading 07", "subheading 60"]
  }
}

Output schemas remain deterministic even when internal components use constrained LLM reasoning.

Error Handling (A Required Part of the Pattern)

A2A provides a consistent, machine-readable structure for failures:

{
  "status": "FAILED",
  "error": {
    "type": "VALIDATION_ERROR",
    "message": "Missing required field: product_description",
    "fields": ["product_description"]
  }
}

Predictable error semantics significantly reduce integration friction.

Why A2A Works Well for Supply Chain Intelligence

✔ 1. It mirrors real multi-step workflows

Developers gain an API abstraction that matches actual execution semantics.

✔ 2. It keeps capabilities independent

Classification logic stays focused on classification; tariff logic stays focused on tariff computation.

✔ 3. It supports long-running jobs cleanly

Whether expanding Tier-10 supplier networks or aggregating due diligence signals, long tasks run safely without timeouts.

✔ 4. It improves transparency and auditability

Each capability has a single responsibility and a deterministic schema.

✔ 5. It avoids monolithic surface areas

Developers integrate only the components they need.

A2A Is Not an Orchestration Framework

A2A does not prescribe:

• how workflows execute internally
• which compute model is used
• whether logic is rule-based, graph-based, or LLM-based
• whether async execution uses queues, workers, or distributed pipelines

A2A simply defines how capabilities expose themselves.

This avoids a common misunderstanding:

A2A is an integration pattern, not a multi-agent runtime.

When A2A Outperforms Traditional APIs

Choose synchronous APIs for:

• small
• atomic
• single-step
• instant operations

Choose A2A for workflows that are:

• multi-step
• long-running
• reasoning-dependent
• requiring deterministic schemas
• naturally modular

Explore the Open Specification

The complete open specification — including lifecycle semantics, schemas, error models, and runnable examples — is available here:

👉 https://github.com/SupplyGraphAI/supplygraph-ai

The repository contains:

• formal manifest definitions
• structured I/O schemas
• consistent error semantics
• example agents
• reference clients
• integration quickstarts

The spec continues to evolve as more capabilities are added across supply chain intelligence:

• multi-tier dependency expansion
• risk propagation modeling
• tariff and compliance automation
• due diligence aggregation
• geographic concentration scoring

If you’ve worked on similar multi-step API designs or long-running workflows, I’d love to hear how you approached them.