Separate agent reasoning from authority
An AI agent can analyze chain data, draft actions, explain risks, and prepare transactions. It should not silently receive permission to upgrade contracts, move assets, change roles, or execute production operations without a policy gate and human approval.
Use typed tools
Tool calls should validate chain ID, contract address, amount, function selector, repository reference, and environment. Allowlisted actions, dry-run modes, and rate limits reduce the chance that natural-language instructions become unsafe execution.
Design wallet flows explicitly
Base applications can combine consumer wallet UX, account-aware product flows, and agent-assisted operations. The interface must distinguish connection, authentication, message signing, transaction signing, and delegated sessions.
Keep audit logs
Record user input, retrieved sources, model version, proposed action, policy decision, approver, tool arguments, transaction hash, and final verification. This creates evidence for incident response and future evaluations.
Start with supervised automation
The safest first release is a copilot that prepares and verifies actions while named operators approve execution. Autonomous execution can be added only after evals, monitoring, and rollback paths prove reliable.