Pre-execution control
AI agent transaction screening
How pre-execution screening checks an AI agent’s proposed transaction before signing or broadcast.
Published by Kondux · Updated July 24, 2026
Direct answer
Direct answer: AI agent transaction screening
AI agent transaction screening is a pre-execution control that decodes a proposed transaction, compares it with the agent’s stated intent and authority, evaluates route and counterparty risk, and returns an allow, block, or review decision before value moves.
What a screening decision should inspect
A transaction can be technically valid and still be economically wrong. Screening should interpret the requested call in context rather than relying on a signature or simulation alone.
- The destination contract, router, token, pool, and counterparty.
- The method, arguments, approvals, transfer amounts, fees, slippage, and expected outputs.
- The user’s request, the agent’s mandate, approved tools, limits, and time window.
- Current risk signals such as honeypot behavior, fake-token patterns, depleted liquidity, or sanctions exposure.
- The expected state change and the evidence needed to explain the verdict.
A practical pre-execution sequence
First, normalize the user or workflow intent into explicit constraints. Second, decode the proposed transaction and simulate the relevant state change. Third, compare the two and apply policy. Finally, produce a machine-readable verdict with human-readable reasons.
A low-risk action that matches intent can continue. A known violation should be blocked. An ambiguous or incomplete action should be held for review rather than treated as safe by default.
Screening versus transaction simulation
Simulation estimates what a transaction will do in a particular state. Screening decides whether that outcome is acceptable for this agent, user, and policy. The two are complementary: simulation supplies evidence, while screening applies intent, authority, and risk rules to that evidence.
Seraph is designed as this policy boundary for agentic commerce, with checks for intent divergence, unsafe approvals, token and route risk, slippage, counterparties, and decision receipts.
Frequently asked questions
Questions about ai agent transaction screening
When should an AI agent transaction be screened?
After the transaction has enough detail to decode and before the wallet signs or the network broadcasts it.
What decisions should a screening system return?
At minimum: allow, block, or review, accompanied by stable reason codes and the evidence used to reach the verdict.
Does simulation guarantee a safe transaction?
No. A transaction may simulate successfully while violating the user’s intent, exceeding the agent’s authority, or interacting with an unacceptable counterparty.