Run an autonomous AI agent with an on-chain FLOP budget

An agent that runs around the clock needs to pay for its own inference without a human approving every call, and without the risk that a bug drains its wallet. Flop's chain was designed for that case. Its spend control doc says an agent "that buys compute continuously needs the spending boundary enforced by the chain rather than by its own good behaviour." This page explains how a builder would set that up when the testnet opens.

Status, November 10, 2026: the agent wallet functions are in the published SDKs and work against the devnet, which earns nothing. The testnet is not live. See the compute use cases hub.

What the use case looks like

Picture a research or monitoring agent that wakes up on a schedule and calls a model a few times. You fund it once, set a ceiling, and let it run. The agent owns a FLOP account with a delegated key. Each job opens an inference session with a miner, the payment sits in escrow until the miner proves the work, and the chain refuses any spend over the limit you set.

Which FLOP mechanisms make it work?

Delegated keys with limits. Flop's technology FAQ says "Agents do not hold the main key: a session or agent key is derived from the main wallet with a spending limit and an expiry." The spend control doc lists the account features: pre-set spend limits, "delegated bounded authority via proxies," multisig, timelocks or vesting, and multi-party escrow pools.

Escrowed sessions. The session lifecycle doc describes four steps: request, match, prove, settle. A request pins five fields: model-weight hash, max latency, compute in FLOPs, a confidentiality flag and the fee. Payment is held in escrow against the proof, and "if the work is wrong, the agent pays nothing and challenges instead."

Disputes. The integration doc says a discretionary challenger posts a 100 FLOP bond, forfeited if the dispute is frivolous, and an upheld fraud verdict "refunds the agent's escrow and slashes the miner in full." The same lifecycle doc says the default SOFT tier's "end-to-end settlement and dispute lane is still being specified." Our proof of useful inference explainer covers the checks.

What the SDK gives you

The SDK reference lists agent wallet functions in the TypeScript @flop-labs/sdk: createAgentWallet, fundAgentWallet, agentTransfer, ownerWithdraw, createSession and executeViaSession. runDirectSession opens a session with a named miner and, per the direct session guide, "verifies every receipt before accepting output or signing payment." The Python flop-sdk has wallet functions for "agent wallets, funding, transfers, withdrawals and session keys." npm listed @flop-labs/sdk 0.5.2 on November 10, newer than the 0.4.0 the docs page names, so read the registry README for the version you install.

Two details from the docs matter for budgets:

• The CLI's --max-flop setting "is the escrow ceiling in FLOP." Cooperative settlement pays the reserved escrow, and "transaction fees are additional."
• The upgrade guide says budget errors such as METERING_REQUIRED should be handled "as unavailable budget authorization. Do not catch them and submit using the old token-rate estimate." It also says "Escrow estimates exclude channel transaction fees."

How you would run it on the testnet

1. Build and test the agent on devnet with the builder quickstart. Use a fresh test wallet, as the quickstart advises.
2. When the testnet opens, give the agent a DID key. The timeline FAQ says that is "the same DID key used on TechnoCore."
3. Claim testnet FLOP from the official faucet. The FAQ says limits apply per identity and the claim period "is not yet decided."
4. Fund the agent's account through a capped session key, not your main key.
5. Let the agent open sessions for its real tasks, and log every channel ID and settlement so you can audit or challenge.

Does an agent with a budget earn the airdrop?

An agent that spends testnet FLOP on settled inference is on Flop's official agent path. The draft airdrop page shares the 1,200,000,000 FLOP agent pool "pro rata to compute purchased in settled sessions," and the timeline FAQ says "the task itself does not matter." A bigger budget doesn't help beyond what the faucet's per-identity limits allow, and splitting spend across many wallets doesn't either: the testnet page scores wallets under common control as one participant.

Faucet grants and holding tokens earn nothing. The allocation is locked at genesis and every 3 FLOP spent on mainnet inference unlocks 1 FLOP, a rule the conversion doc calls "ratified but not yet built." Devnet earns nothing. Our FLOP for AI agents guide and airdrop guide have the detail.

Honest limits

• The SOFT-tier dispute path is still being specified, so challenge mechanics may change.
• Devnet is rate-limited; see devnet limits.
• --new-wallet on devnet saves an unencrypted test wallet. The direct session guide says to "use an encrypted keystore for valuable funds."
• No FLOP token exists yet. Funding a real agent budget is a mainnet question, targeted for Q1 2027.

FAQ

How do I stop a FLOP agent from overspending?

Flop's docs say an agent key is derived from the main wallet with a spending limit and an expiry, and accounts support pre-set spend limits, proxies with bounded authority, multisig, timelocks and escrow. The chain enforces the limit, so a bug in the agent can't exceed it.

What happens if a miner returns bad work to my agent?

The agent can challenge instead of paying. A discretionary challenger posts a 100 FLOP bond, lost if the dispute is frivolous. An upheld fraud verdict refunds the escrow and slashes the miner in full. Flop says the SOFT-tier dispute path is still being specified.

Does a bigger agent budget earn a bigger FLOP airdrop?

The agent pool is shared pro rata to compute purchased in settled testnet sessions, but the faucet has per-identity limits and wallets under common control count as one participant. Holding tokens earns nothing.