Control for agent spend

Give agents a budget, not a blank check.

Set the rules once. Lootrunners handles approvals, spend limits, exceptions, and the audit trail behind every transaction.

Start in sandbox

Start read-only. Add spending authority when you're ready. Revoke it when you need to.

Sample Run · lead researchPlaying sample

11:43Read spend context

11:44DataCo selected$3.80

Total economic cost — AI + services + fees$3.84

Research AgentSandbox

$50 total · $10 automatic · new services need Maya

$31.38remaining authority

Sample data shown as it appears after ledger verification.

AUTOSupport Agent · $0.42 · ticket enrichment · TrustLayerAUTOResearch Agent · $0.08 · verify employer · example-data.comASKContent Agent · $12.00 · stock license · new serviceAUTODev Agent · $1.90 · CI compute burst · CloudRelayBLOCKResearch Agent · $140.00 · subscription · outside authorityAUTOSupport Agent · $0.05 · sandbox trial purchase · BazaarASKFinance Agent · $28.40 · above single-purchase capAUTOResearch Agent · $2.00 · 1,200 records · ScrapeGridAUTODev Agent · $0.65 · sandbox API calls · VectorHubASKContent Agent · $19.00 · first purchase from this service

01 · The problem

A wallet shows where the money went. A Run shows why.

Agents can buy data, call paid APIs, and use compute in seconds. A transaction shows the amount and destination. It does not show what the agent was trying to do, whether the purchase was allowed, or whether the result was worth the money. Lootrunners keeps the payment connected to the work.

Without Lootrunners

19 USDC → 0x8f…

With Lootrunners

Research Agent spent $19 on 2,000 records for Midwest Expansion.
Maya approved the new vendor once. 1,842 records were accepted. Payment reconciled.

See the wallet side for yourself: analyze a public wallet → Read-only, no account.

02 · How it works

Observe. Bound. Run. Learn.

Connect an agent without giving it spending authority. Watch what it asks for, set the boundary, test it with fake money, then move to real spending when you're ready.

01 · OBSERVE

See what it asks for.

Connect an agent in read-only mode. Every request is recorded; nothing is paid.

02 · BOUND

Set the boundary.

Choose what the agent may access, what it may buy, how much it may spend, and when a person must approve.

03 · RUN

Watch the work.

Decisions, tool calls, approvals, payments, and outputs arrive in one timeline.

04 · LEARN

Know what it was worth.

See total economic cost and the outcome it produced.

03 · The Run

The whole job, not just the transaction.

What did it do?

A timeline of the decisions, tools, purchases, and outputs in the Run.

Was it allowed?

Every relevant action shows the Rule, authority, and decision behind it.

What did it cost?

AI compute, paid services, and fees in one economic total.

What did we get?

The outcome attached to the Run, with cost per useful result when possible.

One Run answers the questions that matter — without opening a block explorer.

Lead research · Run #1832Completed

43  Read spend context
44  DataCo selected · $3.80
44  Policy decision · AUTO · SPEND-005
44  Paid · $3.80 USDC · verified
45  Attached output · 247 rows
45  Outcome accepted ✓

What the run cost

AI compute$0.08
Paid services$3.80
Network / fees$0.00
Total economic cost$3.88
247 accepted records$0.0157 / record

04 · Control

A boundary you can actually verify.

Set the boundary in Lootrunners. Verify critical limits independently where the payment rail supports it. Governed decisions keep the rule and evidence behind them.

Control alignment

Declared$25 / requestMandate

Enforced$25 / requestProvider/on-chain control

Observed$12.40Largest payment

✓ Aligned

Shown only after enforcement is verified.

Blocked means the boundary held

20 blocked safely

$332.32 not paid

A policy block is not a loss or an error. Red is reserved for control failures, unexplained movement, failed revocation, or reconciliation problems.

Freeze autonomous spending.

Stop new autonomous payments and revoke supported external permissions. The agent cannot override it.

Resource access is separate from spending authority: connecting a service does not hand it to every agent.

05 · Scout

Ask why. Get the evidence.

Scout explains spend, Runs, Resources, Rules, controls, and outcomes using the records behind them. It can draft a change or exception request. It cannot activate one.

Scout · Lootrunners AIevidence-linked

Why did Research Agent spend $38 yesterday?

12 purchases totaled $38.12.

9 were automatic under Data Research

3 were approved once by you

PeopleData was the largest vendor at $12.40

No purchase exceeded $5

Evidence: 12 spend intents · Run #1832 · Mandate v4 · SPEND-005 v3

View 12 purchasesDraft: lower limit to $5

Evidence-linked.

Answers link back to the purchases, Runs, mandates, and outcomes behind them.

Drafts only.

Scout can propose a lower limit or new rule. A person activates the change.

Capability-limited.

Authority-changing tools are not part of Scout's tool surface.

06 · Developers

One spend interface. Authority stays server-side.

Connect through MCP, REST, or the SDK path that fits your runtime. Lootrunners gives the agent only the capabilities it needs while authority, approvals, Rules, and payment state remain on the server.

GET edge.lootrunners.com/.well-known/lootrunners

{
  "api":  { "baseUrl": "…/api/agent/v1" },
  "mcp":  { "url": "…/mcp" },
  "environment": "sandbox",
  "capabilities": { "spend": true, "runs": true }
}

A deliberately small spend surface.

Current surface: 8 tools.
get_spend_context · search_paid_services · request_spend · execute_paid_request · get_spend_status · submit_outcome · request_exception · list_recent_spend

No increase_limit. No edit_policy. No approve_exception. No reveal_wallet. Authority-changing capabilities are not exposed to the agent.

Generated agent instructions

- Read spend context before planning a paid request.
- Create a spend intent before executing a purchase.
- Respect AUTO / ASK / BLOCK.
- While ASK is pending, poll — do not create duplicates.
- Submit the outcome after a paid service returns.

These instructions guide the client. Authority is enforced by the server.

A spend request, in code

POST /api/agent/v1/spend/intents
{ "mandate_id": "data-research",
  "purpose": "Verify employer and title",
  "amount_usd": "0.08",
  "vendor_domain": "example-data.com" }

// decision.result is AUTO, ASK, or BLOCK

07 · Security

Non-custodial. Non-negotiable.

Signing stays outside the model.

Treasury credentials are not exposed to the agent or model. Execution stays with the configured payment rail, managed signer, or customer-controlled account.

One financial source of truth.

Budgets, approvals, payments, and settlement state are reconciled against canonical Lootrunners records.

People change authority.

Activation, wider limits, live promotion, revocation, and unfreeze remain human actions with step-up authentication. One-time approvals never change future authority.

Access is scoped too.

Connecting a service does not give every agent access to it. Resources can be granted to one agent or one Run.

Put one agent behind a boundary.

Start with fake money and no treasury setup. Watch a Run, learn how the agent behaves, then grant only the access and spending authority it needs.

Start Now Watch the sample Run

Read-only first · non-custodial · freeze anytime