Action turns a reasoning proposal into an interaction with an external system or a structured artifact. It can read a document, call an API, write a file, update a database, send a message, initiate a payment, or deploy code. Tools may be selected at runtime, and their outcomes are uncertain. section 6.1 section 2.2.1
For a state-changing task, the action is a bounded, verifiable change—not merely a tool call. A successful return, a green test, or a model saying “done” is evidence only of what that signal observes. Completion requires task-specific evidence that the intended effect was accepted. section 6.0 section 6.1 section 2.2.1
An action contract connects the intended effect, execution boundary, state basis, commit identity, and acceptance evidence. Governance limits what may execute; reflection compares the result with the intent and can trigger revision or termination. Some effects are irreversible, so controls must operate before execution as well as after it. section 6.9 section 2.2.1
Choose controls for the task’s failure modes rather than adding every pattern:
- Prompt chaining passes named artifacts through checked gates when the sequence is known. section 6.9
- Tool dispatch restricts the eligible capabilities before selection and rechecks the exact call’s authority, provenance, and freshness. Discovery and schema conformance do not grant permission. section 6.9
- Plan-and-execute persists dependencies, progress, uncertain effects, and evidence when restarting after interruption would be unsafe. section 6.9
- A guardrail sandwich places lower-trust execution between protected preconditions and independent postconditions. Containment limits reach; acceptance still requires task-specific evidence. section 6.9
A valid image can still be the wrong result
An agent publishes a map imagery layer and receives a valid PNG, but the intended viewer shows an opaque white rectangle over the basemap. Verification in that viewer checks the requested overlay effect, which the image response alone cannot establish. section 6.0
A paid status can hide a duplicate transfer
A payroll agent makes an authorized payment, then follows an instruction planted in imported data to submit it again. A PAID status does not reveal the duplicate; the action needs a record connecting the authorization to attempts and actual transfers. section 6.0
















