Skip to content

For AI agent builders

Build agents for regulated customers. Prove they follow the rules.

Your customers in finance, healthcare and government need control and proof. TraceStone checks every action your agent takes against their rules before it runs, and gives their auditors that proof.

  • Works with any agent framework
  • Embedded check or gateway
  • Runs in your customer's environment
Example: one AI agent works for three customers, a bank, a hospital and a government agency. Each customer has its own rules. The agent asks to pay S$48,000 to K. Ong. At the bank, payments over S$10,000 need a person, so it waits and is sent after a person approves. At the hospital, the limit is S$50,000, so it is allowed. At the agency, agents may not make payments, so it is blocked and not forwarded. Next the agent asks to read payment credentials, and all three customers block it. Every decision becomes a signed record in that customer's own chain.
One agent. Three customers. Checked by TraceStone
Example
The same action meets each customer's own rules, before it runs.

Unsafe actions stop. Risky ones wait.

Every action is checked against your customer's policy and the user's task before it runs. Nothing outside the rules reaches their systems.

  • Their policy, enforced Your customer's administrators set the main rules for your agent.
  • Task-aware rules Built from each request and formally verified before use.
  • Approval limits Over the limit, an action waits for a person, then runs.
  • Stopped before it runs An action outside the approved task is never forwarded.
TraceStone live requests: a reporting agent asks to send a revenue summary to an audit partner. External destination, approval required: the request is awaiting review and not sent. Three read requests on approved datasets were allowed.
TraceStone action details: approved once and executed. A reviewer approved this specific summary and destination, and the export ran only after that decision. Steps: request received, task rule checked, human approval recorded (CFO approved, summary only), approved summary sent.
TraceStone action details: a data agent asked to read payment-system credentials. Blocked before execution: payment credentials are outside the approved reporting task, and the request was not forwarded.
From the TraceStone product

Hand over proof. Not their data.

Every decision becomes a signed record. Your customer's auditor checks the proof alone, without reading raw logs.

  • Signed, linked records Each record is signed and linked to the one before.
  • Only what the audit needs Selected records go out. Revenue figures, prompts and payloads stay with your customer.
  • Prove without revealing Zero-knowledge proofs show a rule held, without the values.
  • Checked offline Each claim comes back verified, failed or inconclusive.
TraceStone export for the auditor: the file TraceStone-action-records.json, marked "Sent to the auditor", with privacy set to selected metadata only and one signed record of a recorded human approval. Not in the file: revenue figures, prompts and replies, request payloads.
From the TraceStone product

Keep your agent. Add the checks.

Your agent and its framework stay as they are. TraceStone adds a check before each action, and it can stop it.

  • Embedded check Call TraceStone from your agent's own code, before each action.
  • Security gateway Put TraceStone in front of your agent's tool calls.
  • Any agent framework Works with the framework you already build on.
  • Inside their environment Runs on your customer's servers, private cloud or hybrid setup.

Not ready to block? Start with the log collector: it only records. Move to the gateway later.

Your customer's environment Example
Embedded check Inside your agent's code Can block

› READ revenue Allowed · signed record

Security gateway In front of your agent's tool calls Can block

› DELETE customer_records Blocked, not forwarded · signed record

Any agent framework Every decision signed

Your questions. Straight answers.

Building for a regulated customer? Talk to us

Can each customer keep its own rules?

Yes. TraceStone runs inside each customer's environment, and their administrators set the main policies. Task-aware rules come from each user request, the available tools and the workflow. They are checked against that customer's policy and formally verified before use.

What can we hand to our customer's auditor?

Signed, linked records of every decision, and proof that the rules held. Zero-knowledge proofs confirm a rule without revealing the value, for example that a payment stayed below its limit. The auditor checks offline: each claim is verified, failed or inconclusive, and missing evidence never counts as a pass.

How do we add TraceStone to our agent?

Two ways that can block: an embedded check in your agent's own code, or a security gateway in front of its tool calls. A log collector can also record after the fact with almost no changes, but it cannot block. All three work with any agent framework.