Praetor: one system for accountable AI agents
Discover, Govern, Enforce, Detect and Explain work as one. At its core, Praetor checks every action before it executes.
- M-2.1Payments up to S$50,000 to approved counterparties
- M-2.2Above S$50,000: CFO approval within 4 hours
- M-3.1New counterparties: screening and approval first
- TD-7.1Trust T-02 distributions need trustee consent
- G-1Agents never change their own mandate
What each function does
Discover
What AI is acting in our name?Builds a live inventory of the agents and AI-enabled tools operating across the organisation, including AI inside vendor software, with an owner for each.
Agents and AI services that call your systems are identified and registered.
Each agent receives an identity, a scope and an accountable owner before it can act.
- Inventory of agents, tools and the systems they can reach
- Shadow and unregistered AI surfaced for review
- A named, accountable owner for every agent
Govern
What may each agent do?Turns policy, delegation maps, trust deeds and mandates into machine-readable rules: what an agent may execute, what it may only recommend, and what it must never touch.
Mandates are written once, in plain language, and versioned.
Changes to a mandate require approval by people. Agents cannot change their own.
- Mandates with value limits, counterparties and jurisdictions
- Trust-deed and investment-mandate clauses encoded as rules
- Approval routes and review windows for each decision type
Enforce
Is this action within the mandate?Praetor sits between agents and the systems they act on. Every action is checked before it executes: allowed, held for a named person, or blocked.
Checks run before an action reaches a bank, ledger or record.
Nothing held for approval moves until a named person decides.
- Pre-execution checks on every action
- Hold-for-approval with a defined review window
- Hard blocks for actions outside the mandate
Detect
Is something going wrong?Watches behaviour across agents for patterns no single rule would catch: repeated near-limit actions, unusual timing, attempts to widen scope, and signs of manipulated instructions.
Learned detection works alongside explicit rules, so each covers what the other misses.
Alerts go to the agent’s owner and to risk, with the evidence attached.
- Behavioural baselines for each agent
- Alerts on drift, scope creep and attempted mandate changes
- Signals of prompt injection and manipulated inputs
Explain
Who decided, on what basis?Every decision, including every refusal, produces a record a trustee, auditor, board or regulator can read: the action, the rule applied, the outcome and the accountable person.
Records are written at the moment of decision, not reconstructed later.
Explanations are generated from the rule actually applied, not from a model’s guess.
- Plain-language decision records
- Linked, sequential records that show if anything is altered or removed
- Reports for boards, trustees, auditors and regulators
Where Praetor sits
Simplified view. Integrations and deployment are agreed for each client environment.
A record written at the moment of decision
Each record names the rule applied and the accountable person, and links to the one before it, so a gap or alteration is visible.
Fits your environment
Deployed to suit each client’s environment and data obligations, including within the client’s own cloud.
Connects to agents through their tool and API calls, and to systems of record through existing interfaces.
Collects what accountability needs: the action, the rule, the outcome and the people involved. Nothing more.
See Praetor on your own workflows
A confidential briefing for your board, risk or operations team. We walk through one workflow and show what Praetor would allow, hold and block.