# Evidence for the Commerce Agents article

These assets support `agentic-payments-commerce.mdx`. Captured September 7, 2026, from the local Anthropic Commerce Agents demo and evaluation records dated September 6. All business and customer data shown is fictional demo data.

## What each image shows

| Asset | Origin | What it supports |
|---|---|---|
| `shopper-overview.png` | Fresh screenshot of the running retail storefront | Interface context: suggestions, fixture orders and cart |
| `merchant-overview.png` | Fresh screenshot of the running retail merchant portal | Interface context: metrics, alerts, catalog, inventory and assistant |
| `inventory-replay.png` | Recorded retail inventory conversation, repeat 2, turns 1–2, replayed in the frontend | 49-unit proposal and the recorded response to chat approval |
| `camping-replay.png` | Recorded retail planning conversation, repeat 2, turns 1–3, replayed in the frontend | Final $539 cart, including two sleeping bags, and the agent's suggestion to add two more |
| `campaign-replay.png` | Recorded retail campaign conversation, repeat 2, turn 1, replayed in the frontend; supplementary image | Staged campaign preview containing budget, audience and copy |

The original evaluation called the conversation runtime directly and did not use a browser. The replay images are **not original screenshots from those runs** and do not represent a new behavioral test. Playwright intercepts chat requests and supplies the archived stream events to the unchanged frontend. No new model calls or approval-button actions occur. The replay marker at the bottom is an added capture label; the Next.js development indicator is hidden.

The frontend still reads the current mock demo for surrounding dashboard data, fixture orders and product display details. Delivery labels and other date-sensitive presentation can differ from the original run. Personalization-memory reads are replaced with an empty list for capture; archived memory and session details are excluded from the selected evidence. Replay screenshots establish how recorded output is displayed, not the current backend's persisted state.

## Recorded evidence

`recorded-evidence.json` contains the exact six selected prompts and event sequences, plus the recorded cart/change snapshots. Every event sequence was decoded from and compared with its original JSONL trace. Original file paths are relative to the commerce-agents repository, with SHA-256 hashes for verification. All three camping carts are included, with independently calculated subtotals of $649, $539 and $683. The displayed $539 run is identified explicitly as repeat 2.

The camping result includes a qualification: the initial plan suggested sharing bedding and said two more sleeping bags would add $178. The final response suggested adding them. The article therefore describes incomplete coverage of the original four-person request rather than an entirely unnoticed quantity mismatch.

`campaign-field-comparison.json` is a derived projection: selected submitted fields, followed by every returned change-item field name. Those wrapper keys are editorial labels, not the tool's response schema. The complete selected events remain in `recorded-evidence.json`.

## Source code

Repository: https://github.com/anthropics/commerce-agents

Pinned upstream commit: `fd4d59224ab96b43c6dc6888207c67b3bd5a24cf`.

The local checkout contains pre-existing OpenRouter, trace-recording and runtime changes. These screenshots use that checkout, not a clean upstream install. No commerce-agent source changes were made for this capture. The two excerpts in the article were compared byte-for-byte, apart from common leading indentation, against the pinned upstream files:

- `merchant-agent/core/merchant_agent/gates.py`, lines 207–215: host approval branch and final return from `check_apply_change`. Earlier provenance and guardrail checks are omitted from the excerpt.
- `examples/demo_common/merchant_fixtures.py`, lines 309–311: audience/copy loop inside `stage_campaign`. The preceding budget construction is explained in the article and visible in the linked full function.

Anthropic source excerpts: Copyright 2026 Anthropic PBC, Apache License 2.0. See `APACHE-2.0.txt`. Snippets are excerpts from functions, not standalone programs.

The supporting JSON and notes are selected development evidence, not an independently replicated benchmark.