Monitoring tells you what happened. Agent Shield governs what is allowed to happen. The distinction is the difference between an audit trail and a governance framework.
Every agent action passes through multiple independent evaluation layers, and any one of them can stop it. An agent cannot skip the process for an action routed through Agent Shield.
The engine evaluates in real time. The result is a single verdict for every action, every time — approved, held for a human, or blocked — with a signed receipt.
AI agents are not software to be secured. They are operational entities — they initiate transactions, make decisions, and create operational exposure. They require the same governance infrastructure as any operational entity within defined authority limits.
The governance framework defines the operational boundaries within which agents operate. Operational controls, authenticated audit trails, and a kill switch that works at every autonomy level.
One published standard and the work built around it frame the platform:
When agents from different organisations transact, governance must extend across boundaries. Agent Shield is building the trust infrastructure for AI agent networks — the equivalent of what established trust frameworks provide for traditional commerce. The features below are in development: the code exists and is switched off by default.
An agent carries a verifiable governance certificate, so a counterparty can confirm its compliance status before transacting.
Governance evaluation designed to extend across organisational boundaries, so a cross-agent transaction proceeds only when both sides’ governance agrees.
Cryptographically linked audit trails across organisations, so both parties can verify a cross-boundary interaction.
Every agent action measured for operational impact in real time. If it cannot be measured, it cannot be governed.
A ten-level model from fully supervised to fully autonomous, with promotion on evidence. Enforcement on live decisions is not switched on yet.
Replay past actions against a changed mandate or authority schedule, and compare mandates, before the change goes live.
Every decision carries an Ed25519-signed receipt you can verify offline. Records are hash-chained, so tampering is evident.
EU AI Act, ISO 42001, GDPR Article 22 and FCA four-eyes controls mapped, with evidence packs built from the decision record. The mappings are self-assessed.
Per-agent baselines and risk weights adapt from review outcomes and overrides, so the system learns from the decisions your team makes.
AGS v3.1 is an open specification published by AgentGoverning. Any organisation can read it and build to it. Agent Shield is the reference implementation. This is how infrastructure standards are built — openly specified, proprietarily implemented.
Many governance tools rely on the agent following its instructions. Agent Shield’s constraints sit outside the agent. They are structural: no agent instruction, no business justification, and no level of agent autonomy can override them.
A dashboard shows you what is happening. A Compliance Certificate proves what has been verified. Issued per sandbox and period, signed, with a cryptographically authenticated audit trail behind it. Hand it to a regulator or present it to a board as governance evidence.
Agents earn expanded authority through a graduated pathway. Promotion requires evidence. Enforcing the model on live decisions is not switched on yet; today an operator promotes and demotes. There is no on/off switch — there is a governed progression.
The system builds an increasingly precise understanding of every governed agent. Its baselines and risk weights adapt from review outcomes and overrides. Static rule checks miss multi-step behaviour, so Agent Shield also looks at the sequence.
Agent Shield is not a product you trial. It is infrastructure you govern with. Request a governance assessment with your team.
The specification is open. The reference implementation is proprietary. This is how standards are built.