Request access to the Stablemode A2A payments dataset

Free for research and evaluation under CC BY-NC 4.0. Access is approved automatically. Commercial use needs a separate license from Stablemode.

Log in or Sign Up to review the conditions and access this dataset content.

Stablemode A2A Payments (Sandbox)

3,000 agent-to-agent payment attempts. Each one was checked against the paying agent's owner policy and then settled, declined or rejected by a running Stablemode instance.

Every row was produced by sending a real API call to the live Stablemode sandbox and recording what the app returned. The outcome, decline reason, x402 payment requirements and transaction id all come from the API response. None of them were written or labelled afterwards.

🧪 Sandbox data. The agents, prices and resources are a simulated economy. No real funds moved, and nothing was sent to a payment network. This is not real payment or customer data.

At a glance

Payment attempts 3,000
Settled 1,588 (52.9%)
Declined by owner policy 1,248 (41.6%)
Rejected by the protocol (wrong payer, already paid, expired, …) 164 (5.5%)
Settled volume (sandbox USD) $3,176.49
Agents / fleets 13 / 3
x402-style payment requests 2,098
Channels runtime: 2,710, console: 290

Why it's useful

  • Policy-aware payment agents: given a payment request and the payer's policy, predict or explain whether it should settle (outcome, decline_category).
  • Testing agent wallets and guardrails: realistic edge cases, such as over-cap amounts, unapproved counterparties, paused agents, double-pay attempts, expired requests and paying a request meant for another agent.
  • x402 integrations: every request carries x402-style payment requirements (scheme: exact, USDC atomic units, payTo, resource, expiry).

Configs

events (default): one row per payment attempt

Column Meaning
scenario What the agent tried (see the table below)
channel runtime (the agent's own token) or console (owner-simulated in the UI)
payer_*, payee_* Agent ids, names, fleets and roles
amount, currency, memo, resource What was being paid for
payer_status, payer_daily_budget, payer_per_payment_cap, payer_approved_payees, payer_policy_version The payer's owner policy at the moment of the attempt
payer_spent_today_before Settled spend earlier the same UTC day
payee_is_approved Whether the payee was on the payer's approved list
payment_request_id, x402_requirements For request-based payments: the request and its x402-style requirements (JSON string)
outcome settled, declined (owner policy) or rejected (protocol or HTTP error)
decline_reason, decline_category The exact message from the API, plus a normalized category
http_status, transaction_id, transaction_status Raw API result

payment_requests

Every x402-style request issued during the run, with the last status it was seen in (open, paid, expired, cancelled).

agents

The 13 agents, their fleets and roles, and their final owner policies. Owners changed policies during the run (policy_version), and one agent at a time was paused for a while.

Outcomes by category

decline_category Events Share
daily_budget_exceeded 531 17.7%
per_payment_cap_exceeded 447 14.9%
payee_not_approved 230 7.7%
wrong_payer 44 1.5%
already_paid 37 1.2%
payer_not_active 36 1.2%
request_cancelled 31 1.0%
self_payment 27 0.9%
request_expired 25 0.8%
payee_paused 4 0.1%

Scenarios

scenario Events Share
x402_request_pay 1,747 58.2%
direct_transfer 626 20.9%
x402_over_cap 129 4.3%
x402_unapproved_payee 122 4.1%
transfer_unapproved_payee 112 3.7%
transfer_over_cap 100 3.3%
x402_wrong_payer 44 1.5%
x402_double_pay 37 1.2%
x402_cancelled_request 31 1.0%
self_payment 27 0.9%
x402_expired_request 25 0.8%

How it was made

scripts/generate.py builds three fleets of agents with different owner policies through the public API. It then runs a weighted mix of scenarios against the live sandbox at a gentle pace: request-then-pay, direct transfers, unapproved payees, over-cap amounts, wrong payer, double pay, cancelled and expired requests, and self-payment. Along the way, owners periodically adjust budgets and approved payees and pause or resume agents. Each attempt records the policy snapshot the payer had at that moment.

Limitations

  • Simulated economy: prices, catalog items and agent behaviour are scripted. Use it to learn policy behaviour, not market behaviour.
  • All activity falls within a single run of a few minutes, so daily budgets apply to the run as a whole.
  • Policy rules are deterministic, so a model can learn them exactly. The data is most useful for evaluating agents and guardrails, and for explanation tasks.

License and commercial use

Released under CC BY-NC 4.0, free for research, education and evaluation with attribution. Commercial use (training or evaluating models or products you sell, or redistributing the data) needs a separate license from Stablemode. Open a discussion in this dataset's Community tab, or contact @sriram1983007-dev.

Credits

Generated against the Stablemode sandbox. Stablemode was founded by Amin (@aminheid). Agent-to-agent payments and this dataset are by Sriram Ramakrishnan (@sriram1983007-dev).

Downloads last month
97

Models trained or fine-tuned on stablemode/a2a-payments-sandbox

Collection including stablemode/a2a-payments-sandbox