a2a — When Both Sides Are Agents

An open benchmark for agentic e-commerce. We simulate full commercial transactions where both the buyer and the seller are AI agents, measure what happens to funnels, negotiations, and payments, and publish every result.

Showcase: a2a.chernov.io • Article: chernov.io/writing/when-both-sides-are-agents • License (code): Apache-2.0 • License (content): CC-BY-4.0

---

The question

E-commerce funnels were built for human attention — urgency badges, countdown timers, social proof, friction-optimised paywalls. When agents do the buying and the selling, none of that matters anymore. But what does matter? How should a seller agent design its onboarding for an audience that doesn't click? How should a buyer agent assemble its context? And when a transaction happens, can either side explain why?

This repo is an attempt to answer those questions experimentally, with versioned code and reproducible results.

---

Quick start

# 1. Install uv if you don't have it
curl -Lsf https://astral.sh/uv/install.sh | sh

# 2. Clone and enter
git clone https://github.com/chernov-consulting/a2a && cd a2a

# 3. Copy config templates
cp env.yaml.example env.yaml   # edit with your API keys
cp .env.example .env           # add provider keys here

# 4. Install dependencies
uv sync

# 5. Run a quick smoke test (uses a mock LLM, no API keys needed)
uv run pytest tests/ -x -q

# 6. Run the v0.1 SaaS subscription experiment
uv run a2a sim --experiment experiments/2026-05-saas-subscription/config.yaml

# 7. Generate the HTML report
uv run a2a report --experiment experiments/2026-05-saas-subscription/

# 8. Open the report
open experiments/2026-05-saas-subscription/report.html

---

What gets simulated

Inputs (product URL/PDF + buyer brief)
  │
  ├─► Catalog fetcher → structured product schema
  │
  ├─► Seller agent   → funnel variant (API contract | rich schema | persuasion page)
  │
  └─► Buyer agent    → mandate (terse JSON | full transcript | annotated brief)
            │
            ▼
        A2A negotiation (≤3 turns)
            │
            ▼
        Decision: buy | walk | escalate_to_human | timeout
            │
            ▼
        L3 mock settlement  ←── no real money, ever
            │
            ▼
        JSONL ledger (tokens, latency, $cost, steps, outcome, rationale)
            │
            ▼
        HTML report + showcase site

3 buyer archetypes × 3 seller funnel variants × 2 product categories = 18 cells, 30 dyads each.

---

Experiments

Slug	Product	Status
2026-05-saas-subscription	B2B SaaS observability tier	v0.1
2026-05-physical-good	Single-bottle wine subscription	v0.1

---

Protocol samples

Runnable, versioned mini-examples in src/a2a/protocols/:

Protocol	Layer	Status
AP2	Authorization	v0.1
MCP	Intent	v0.1
A2A	Intent	v0.1
x402	Intent + Authorization	v0.1
UCP	Intent	v0.1

---

Citing this work

Chernov, A. (2026). When Both Sides Are Agents: A Versioned Open Benchmark
for Agentic E-Commerce. Chernov Consulting OÜ.
https://github.com/chernov-consulting/a2a

Full BibTeX in NOTICE.

---

Contributing

Issues and PRs welcome. See AGENTS.md for code conventions, commit format, and security rules. All contributors agree to license their contributions under Apache-2.0 (code) / CC-BY-4.0 (content).