Governed AI-agent actions across business systems

An AI Agent can propose a useful business-system change without receiving ambient authority to perform it. PanSphaira's released local reference path demonstrates that separation with fictional CRM and ERP records.

The bounded path

1. An Owner asks for an outcome and the Agent forms a typed proposal.
2. The Gateway evaluates the declared operation against the current authority profile and Policy.
3. An effect that requires consent becomes an Owner-visible approval, bound to the exact proposal and authoritative pre-effect snapshot.
4. The Broker executes only the admitted effect.
5. Provider readback and a digest-bound receipt establish the local synthetic result. Transport acceptance alone never becomes success.

The SAFE_GUIDED proof binds these statements to machine-checked claims and negative probes. The Architecture describes the local trust boundaries.

Who this is for

This proof is useful to engineers evaluating patterns for governed Agent effects, approval integrity, fail-closed execution, and evidence-oriented business-system integration. Run it when a deterministic local reference is useful before live-system or tenant work begins.

Use a conventional workflow engine when the work is fully predetermined and does not need Agent proposals. Use a policy engine when policy evaluation is the only missing component. Use an Agent framework when experimentation is the priority and no effect authority is present. PanSphaira does not claim to replace those systems or outperform them.

What this does not prove

The current path is one pinned Linux/Docker fixture using fictional data. It is not production IAM, a security certification, a live CRM/ERP integration, a general approval service, or evidence that arbitrary Agent/tool/network paths are completely mediated. Review the full known limitations before reproducing the example.

Next action

Inspect the specific CRM-to-ERP approval and readback path, use the Quickstart for the complete local demo, or choose a smaller deterministic command from the examples gallery.