From request to result
1
The agent requests an outcome
The request identifies a paid capability or financial action and its purpose.
2
The request becomes bound semantics
Agent identity, funding source, provider, method, resource, asset, amount, recipient, route, terms, and expiry must be explicit. Missing or ambiguous details produce a question or denial, not a guess.
3
Policy returns allow, ask, or deny
Deterministic rules check capability, provider and venue allowlists, per-call and cumulative budgets, reusable scope, and approval thresholds.
4
Authority and budget are reserved
A permitted request binds the exact policy version, route, debit ceilings, expiry, and single-use execution authority. An
ask decision waits for independent approval.5
The action is prepared and revalidated
Supported adapters prepare and simulate typed steps. Settlement, resource delivery, transaction submission, and finality remain distinct states.
6
The outcome is accounted for
Denied, purchased, delivered, prepared, submitted, failed, and completed outcomes enter one readable audit record.
Why the model is not the control plane
Language models are useful for interpreting requests and explaining results, but their output can be incomplete, inaccurate, stale, or prompt-injected. please does not treat model prose, tool output, a previous approval, or a successful simulation as executable financial authority. The control plane and planning engine instead work with versioned, typed artifacts and fail closed when required identity, policy, budget, configuration, or evidence is unavailable.What a plan contains
A useful plan can include:- Agent, funding source, and purpose
- Policy decision, version, and approval requirement
- Budget reservation and debit ceilings
- The canonical action and material parameters
- Human-readable steps
- Expected asset deltas
- Network, protocol, and destination information
- Slippage or fee limits when applicable
- Warnings and unsupported conditions
- Simulation mode, state reference, and expiry
The exact fields depend on the action and the service configuration. Always use
discover_actions before a public MCP simulation.