A2A

From Email Chains to Agent Delegation: B2B RFQ Automation with A2A and Hermes Agent

By IdeaBosqueLast updated: July 24, 2026

Key takeaways

Gartner predicts 90% of B2B buying will be AI-agent-intermediated by 2028, channeling $15 trillion through agent exchanges — the two-year runway means agent-readiness is now a commerce architecture mandate, not a future research item.
Manual RFQ processing takes 3.4 weeks per contract cycle and 30 minutes per quote on average — Forrester and Aberdeen research shows automation can compress this to minutes, and one distributor reported a 6× speed increase (30 minutes to 5 minutes per quote).
AI-enabled procurement workflows report 30% reductions in manual work and up to 45% cost reductions — Automation Anywhere's 2026 procurement data shows the gains are real, not projected.
The Agent2Agent Protocol surpassed 150 organizations in its first year — Linux Foundation governance, production deployments across supply chain, financial services, and IT operations, with deep integration across Google, Microsoft, and AWS platforms.
Hermes Agent from Nous Research ships with 60+ built-in tools and native MCP support — an agent that can be pointed at RFQ, catalog, and pricing MCP modules without custom integration glue, and can speak A2A through a bridge layer without rewriting its internals.

The problem: a 3-day RFQ cycle that does not scale

A mid-market B2B distributor receives 200 RFQs a week. Each arrives as an email with a spreadsheet or PDF — line items, quantities, requested delivery dates, sometimes a customer account number. The quoting team opens each one, looks up each item in the catalog, checks current inventory levels, matches the customer to the right pricing tier, pulls cancellation policies from the supplier's terms, drafts a quote, and sends it back. If the RFQ involves multiple suppliers — a bundle of components from three vendors — the process multiplies. Someone coordinates with each supplier by phone or email, waits for responses, reconciles availability, and assembles the combined quote.

The result is a 3-day cycle. Forrester and Aberdeen research shows that manual contract approval alone takes 3.4 weeks on average. One distributor measured 30 minutes per quote before automation; after, 5 minutes — a 6× speed increase. The 45% cost reduction and 30% reduction in manual work reported across AI-enabled procurement workflows are not theoretical projections. They are measured outcomes from teams that replaced the email-and-spreadsheet cycle with structured automation.

The problem is not that the quoting team lacks effort. The problem is that the workflow is built around manual lookups against disconnected systems. The catalog lives in BigCommerce or a PIM. Inventory lives in NetSuite. Pricing tiers live in a spreadsheet or a CRM. Cancellation policies live in supplier PDFs. Each system has its own API, its own authentication, its own data model. A human bridges them by reading screens and typing values. An agent can bridge them by calling tools — but only if those tools are exposed in a way the agent can use.

Manual workflow versus agent-orchestrated workflow — what changes when MCP tools, A2A delegation, and Hermes Agent replace email-and-spreadsheet quoting:

The agent-orchestrated solution: MCP tools, A2A delegation, Hermes orchestration

The architecture has three layers, each addressing a specific bottleneck in the manual workflow:

Layer 1: MCP modules give the agent hands

The Model Context Protocol exposes each business system as a set of typed tools an agent can call. Instead of a human reading a NetSuite screen, an MCP module wraps the NetSuite SuiteTalk API behind tools like get_inventory, check_availability, get_price_tiers. Instead of a human browsing a BigCommerce catalog, a module exposes search_items, get_provider_items. The RFQ engine itself — the system that manages request lifecycle, quote creation, availability holds, and cancellation policy snapshots — exposes 38 registered MCP tools across 11 domain mixins: request submission, item search, availability holds, quote management, pricing calculation, installment scheduling, bundle assembly, cancellation policy retrieval, file attachments, segment contacts, and catalog inquiry via knowledge graph search.

Each tool has a typed schema. The agent does not guess what arguments to pass — it receives a JSON schema with field names, types, and descriptions. This is the difference between an agent that drafts a plausible-looking quote and an agent that calls acquire_availability_hold with a real provider_item_uuid and gets back a hold_token with a 15-minute TTL. The typed schema is what makes the agent's output verifiable rather than decorative.

Layer 2: A2A gives the agent colleagues

The Agent2Agent Protocol — now under Linux Foundation governance with 150+ supporting organizations — defines how agents discover each other, delegate tasks, stream progress, and report completion. In the RFQ workflow, A2A solves a specific problem: the quoting agent does not need to be a single monolith that knows everything. It can delegate.

A buyer-side agent receives an RFQ, extracts line items, and delegates the catalog lookup to a supplier agent that owns the product database. The supplier agent searches the knowledge graph, returns matching items with availability and batch information. The quoting agent delegates availability verification to an inventory agent that owns the NetSuite connection. The inventory agent acquires holds. The quoting agent delegates pricing tier resolution to a pricing agent that owns the customer-segment logic. Each agent owns one system and one domain. A2A is the protocol that lets them hand work to each other, track it, and receive results — without any agent needing to know how the others are built.

The docker-a2a-hermes-agent-gateway project demonstrates this pattern as a working, open-source deployment. A gateway handles authentication, per-customer routing, and rate limiting. A bridge layer translates A2A JSON-RPC calls into the agent's native API. The agent itself — Hermes Agent — does the reasoning and tool calling. PostgreSQL Row-Level Security enforces tenant isolation at the database level, so a coding mistake cannot leak one customer's RFQ data to another.

Layer 3: Hermes Agent orchestrates the workflow

Hermes Agent from Nous Research is the orchestration layer. It ships with 60+ built-in tools, native MCP client support, and a skill system that lets it learn from experience. For the RFQ workflow, the agent is configured with:

MCP modules that expose the catalog, inventory, pricing, and RFQ engine as callable tools
A2A bridge that lets it delegate to and receive delegation from other agents — supplier agents, inventory agents, pricing agents — without any of them sharing a framework
Human-in-the-loop approval gates that pause the workflow when a quote exceeds a margin threshold, requires a custom discount, or touches a regulated product category. The agent raises a "waiting for approval" state that carries across the entire agent chain, even across framework boundaries. A human approves or declines, and the work resumes.
Audit logging — every tool call is recorded with status, duration, and result. The quoting team can trace exactly which tool produced which value in the final quote, which is what a compliance reviewer needs.

The workflow runs as follows. An RFQ arrives. The agent parses the line items, calls the catalog MCP module to match each item, delegates availability checks to the inventory agent via A2A, calls the pricing module to resolve the customer's tier, acquires availability holds, retrieves cancellation policies, assembles the quote, and — if the margin or discount falls outside pre-set guardrails — pauses for human approval. If approved, it sends the quote. The entire cycle takes hours, not days.

The outcome: what changes for the business

The measurable changes map directly to the bottlenecks identified above:

Metric	Manual workflow	Agent-orchestrated
Quote turnaround	3 days per cycle	4 hours per cycle
Manual work per quote	30 minutes	5 minutes (agent does the rest)
Availability hold accuracy	Manual lookup, errors common	Atomic hold with 15-min TTL, no oversell
Cancellation policy	Copy-paste from PDF, stale risk	Snapshot frozen at quote time
Pricing tier errors	Customer matched to wrong tier	Typed schema + segment resolution
Multi-supplier coordination	Phone and email, 2-3 days	A2A delegation, parallel, minutes
Audit trail	Email threads, incomplete	Every tool call logged with status and duration
Human touchpoints	Every step	Only margin-guardrail and regulated approvals

The 3-day-to-4-hour compression is the headline number. But the operational changes underneath it matter more. Availability holds that do not oversell. Cancellation policies that are snapshot-frozen, not stale copy-paste. Pricing tiers resolved by typed schema, not human memory. Multi-supplier coordination that runs in parallel through A2A delegation instead of serial phone calls. And an audit trail that records every tool call — which is what a Head of Engineering needs when the CFO asks why a quote was priced the way it was.

The 30% reduction in manual work and 45% cost reduction reported by Automation Anywhere across AI-enabled procurement workflows are the industry-level evidence. The per-distributor numbers — 6× quote speed, 3-day-to-4-hour cycle — are what that looks like when the workflow is a B2B RFQ process rather than a generic procurement function.

The broader context: why this matters now

Gartner's prediction that 90% of B2B buying will be AI-agent-intermediated by 2028 — channeling $15 trillion through agent exchanges — is not a reason to panic-build. It is a reason to understand the architecture decision in front of you. The question is not whether agents will participate in B2B commerce. The question is whether your systems expose their capabilities in a way an agent can use — typed tools, standard protocols, auditable processes — or whether they remain behind human-readable screens that an agent cannot reach.

MCP gives agents hands. A2A gives agents colleagues. Hermes Agent gives the workflow an orchestrator that can use both without locking you into a single vendor's framework. The bridge pattern means you adopt the standards without rewriting the agents you already have. The tenant isolation means one deployment serves multiple business units without data leakage. The human approval gates mean autonomy does not mean uncontrolled.

For a Head of Engineering or VP of Operations running a quoting team that processes 200 RFQs a week through email and spreadsheets, the decision is not whether to automate. The 3-day cycle is already losing deals to faster competitors — the distributor that quotes in 4 hours wins the business that the 3-day distributor loses. The decision is which architecture to build on: one that locks you into a single vendor's proprietary agent runtime, or one built on open protocols (MCP, A2A) with a swappable orchestration layer (Hermes Agent) that you can replace without re-doing the integrations.