Guide · Controlling AI agent actions
How to control and prove what your AI agents do
AI agents now issue refunds, move money and change customer accounts. To trust them with that work, a business needs limits enforced outside the agent, a person in the loop for risky actions, and a record of every action that others can check. Tridris is the control and proof layer for AI agent actions: it sits between your agents and the systems they act on, checks each action before it runs, and proves what happened with records your auditors, insurers and customers can rely on.
Prefer email?info@tridris.com
What to look for in a tool that controls AI agent actions
Limits enforced outside the agent
Instructions in a prompt can be ignored or overridden. Per-action limits, daily budgets and approval thresholds have to be checked before each action runs, by something the agent cannot talk its way past.
A person for the risky actions
Routine actions should run straight away; large refunds, payouts and identity changes should wait for an authorized person, who sees why the action was held.
Every agent in one place
Teams build agents on different models, frameworks and agent builders. The controls should cover all of them, not one at a time.
A record others can check
Logs written by the system they describe prove little in a dispute or an audit. Each action needs a record that customers, auditors and insurers can verify themselves.
A way to stop everything
When something goes wrong, you need to pause every agent, or one group, at once.
A safe way to start
You should be able to see what your limits would have done before you enforce them, and keep the credentials that move money in your own systems.
Why teams choose Tridris
Proof your auditors can rely on
Records are signed by Tridris and countersigned by a separate sealing service. Evidence packages carry an offline verifier, so checking the proof does not depend on access to Tridris or your systems. Each record shows the state before and after the change, and confirms the change in your system of record. Records are also timestamped by DigiCert, an outside timestamp authority, so no one, not even Tridris, can change, delete or backdate a record without it showing.
Every agent, whatever it was built with
Tridris finds the agents your business already runs, connects to any agent through MCP or its API, and governs them with limits set for groups, the way you would set them for teams of employees.
Business actions, not prompts
Tridris decides whether this refund, for this amount, on this order, today, runs, waits for a person or is blocked.
Check · Trace · Seal in one product
Built for the people who own the risk, such as support, finance, operations and risk teams, not only for engineers.
What Tridris does today
Built-in checks for six customer support and e-commerce actions, plus any action you define and any MCP tool your agents reach through Tridris. Ready-made checks for other industries: talk to us.
Agent discovery
Connect where your agents are built (n8n today) and Tridris lists every agent it finds. Nothing is governed until you choose to.
Connect any agent
An MCP server and an API: any agent that can call a tool can ask Tridris first, whatever it was built with.
Your own actions and tools
Define any action your agents take, and route their MCP tools (a database, a deployment server) through Tridris: every call is allowed, held for approval or blocked by your rules, and the agent never holds the tool's credential.
Groups and a company baseline
Set limits once for a group of agents, like a team of employees. Agents can follow the group or have their own limits.
Emergency stop
Pause every agent, or one group, in one click. Every request is refused and recorded until you resume.
Before and after proof
Each completed action records the state of the affected item just before and just after the change (as fingerprints), and a separate record confirms the change in the system of record, or flags that it is missing.
Evidence for the reader
Export evidence for a customer dispute, an audit, an insurance claim or a board report. Every package includes an offline verifier that runs in any browser, with no network.
Public verifier
An auditor, insurer or customer can drop an evidence package on Tridris's verify page and check it in their own browser. Nothing is uploaded.
Outside timestamps
About every 10 minutes, DigiCert, an outside timestamp authority (RFC 3161), timestamps the latest records. The offline verifier checks the timestamp with no network.
Unidentified activity
Store check (reconciliation) flags actions in your systems that did not go through Tridris.
Simulate
Replay recent actions against your limits to see what would have been held or blocked.
Decision-only
Tridris decides and signs; your own system carries out the action with your credentials. The recommended start for design partners.
How Tridris compares
Tridris works alongside the tools you already use. Here is how it differs from each kind.
Model monitoring and guardrails
Watch what a model says: prompts, responses, evaluations. Tridris governs what an agent does to the business, and proves it.
Identity and access tools
Decide who an agent is and what it can sign in to. Tridris decides whether a specific action runs now, given its amount, its customer and today's spending.
Agent gateways and runtime authorization
Sit between agents and their tools, and allow or block each call by policy. Tridris also checks the business details of each action, such as the amount, the customer and today's spending, read from your systems rather than from the agent. It sends risky actions to a person and seals a record that can be checked outside Tridris.
Logs
Are written and kept by the system they describe. Tridris records are countersigned by a separate sealing service, timestamped by DigiCert, an outside authority, and your auditors, insurers and customers can check them for themselves, offline.
Questions people ask
- What is Tridris?
- Tridris is the control and proof layer for AI agent actions. It sits between your agents and the systems they act on, checks each action before it runs, and proves what happened with records your auditors, insurers and customers can rely on.
- How do I stop an AI agent from issuing large refunds or payouts?
- Set per-action limits, daily budgets and an approval threshold for the agent in Tridris. Actions within your limits run straight away, larger ones wait for an authorized person, and anything above your maximum is blocked. Every attempt, including blocked ones, ends in a verifiable record.
- How do I prove what an AI agent did?
- Every action Tridris handles is sealed into a verifiable record: signed by Tridris, countersigned by a separate sealing service, and linked to the record before it. About every 10 minutes, DigiCert, an outside timestamp authority, timestamps the latest records, so no one, not even Tridris, can change, delete or backdate a record without it showing. You can export an evidence package for a dispute, an audit, an insurance claim or the board, and your auditor or insurer can check it with the offline verifier inside. Any edit to a record is detected.
- So I have to trust you?
- Only that we record the truth at the moment an action happens. After that, every record is timestamped by DigiCert, an outside authority, so no one, not even us, can change, delete or backdate it without it showing, and you can check that offline yourself. We're also building an option for enterprise partners to hold a key themselves, so they don't need to trust us for that either.
- Can I approve AI agent actions from my phone?
- Yes. Approvers decide in the Tridris dashboard or from a link on their phone, and see why the action was held. Passkeys, such as Face ID or a fingerprint, are optional for each approver.
- Does Tridris work with agents built on Claude, OpenAI, Gemini, Grok, n8n or our own code?
- Tridris connects to any agent through an MCP server and an API, so any agent that can call a tool can ask Tridris before it acts. Agent discovery works with n8n today, with more agent builders added with design partners. Tridris is not affiliated with or endorsed by these companies.
- Do I have to give Tridris my payment credentials?
- No. In Decision-only mode, Tridris decides and signs, and your own system carries out the action with your credentials. Tridris never stores card or bank details. In Enforce mode, Tridris holds the store and payment credentials you provide, keeps them only in its service and never gives them to agents.
- Can I try it without changing anything?
- Yes. Simulate replays your recent actions against the limits you set and shows what would have been held or blocked, without changing anything in your systems.
- How is Tridris different from AI monitoring or guardrail tools?
- Guardrail and monitoring tools watch what an AI says. Identity tools control what an agent can sign in to. Tridris decides whether each action runs, for how much and with whose approval, and proves what happened. It works alongside the tools you already use.
- Which actions does Tridris govern today?
- Live today: built-in checks for six customer support and e-commerce actions (refunds, store credit, order cancellations, shipping address changes, account email changes and discount codes), any other action you define in Tridris, and any tool your agents use through the Tridris MCP gateway. Ready-made checks for other industries: talk to us.
- How do I get started?
- Tridris is working with a select group of design partners. Book a live demo or apply through the design partner form, or email info@tridris.com.
Put AI agents to work with confidence.
Book a live demo, or apply to become a design partner and help shape the actions Tridris covers next.
Prefer email?info@tridris.com