
Every control switched on blind is a guess
Security teams want controls on their AI agents: refuse unknown agents, mask customer data, hold budgets, limit which credentials an agent may use. Switch one on without evidence and an agent the business depends on can stop mid-task. Leave it off and nobody can say what it would have caught.
- IBM found that security incidents involving unapproved ("shadow") AI more than doubled, to 43% from 20%, and 92% of organizations with an AI-related breach lacked proper AI access controls.¹
- Good practice already asks for the evidence first: NIST SP 800-53 control CM-4 says to "analyze changes to the system to determine potential security and privacy impacts prior to change implementation."²
Both ways, the security team is guessing. Watch mode replaces the guess with your own traffic.
The one idea: every control has a watch switch
Argent Strategic Intelligence sits between your AI agents and the models and tools they use. Twelve of its controls each have their own switch, WATCH or ENFORCE, set per tenant.
- In WATCH, a control runs its full check on every request that goes through the Argent gateway, sends the request on unchanged, and writes what it would have refused or changed onto the signed record.
- In ENFORCE, it refuses and changes requests as configured.
- A new tenant starts with every control in WATCH. One "Watch all nodes" switch holds every control in watch at once and keeps each control's own setting underneath.
- The twelve: Identity Governance, Policy (PDP, the policy engine), Credential Broker, DLP (Data Loss Prevention), Cost and Budget, MCP Gate (agent tool calls), A2A Gate (agent-to-agent calls), RAG Authz (which documents an agent may retrieve), AI Sentinel (prompt-attack screening), Untrusted Fencing, Add-ons and NHI behaviour (how each non-human identity behaves).
How it works, step by step
One thread, from a single request to a signed switch.
- An agent sends a request. A branch assistant asks a model to summarize a loan file that holds a customer's Social Security number. DLP is in WATCH.
- The check runs in full. DLP finds the number exactly as it would in enforce. The model receives the agent's original request, and the agent sees no added header and no changed answer.
- The record says what it would have done. The signed audit row names the kind of data and the count, never the value. On the Live traffic page the row reads
ALLOWwith "DLP detected SSN x1 and sent the request unchanged." - The Rollout page counts it. For each control: how many requests it would have refused or changed, which agents it touched, and sample events.
- Admins judge the samples. Each sample is marked RIGHT or WRONG, and each label is a signed admin event. After five labels the page shows how often that control was wrong.
- Flip one control. The dialog states the impact first, for example "Enforcing DLP would have refused 14 requests from 3 agents in the last 7 days." A fresh passkey check signs the change and it applies live. DLP now enforces; the other eleven controls keep watching.
What the operator sees
The rollout work lives on one console page, ROLLOUT, readable by SuperAdmins, TenantAdmins and Auditors.
| Screen | What it shows |
|---|---|
| Register | Each control's would-refuse and would-change counts, trend, agents hit, false positives, who set it, and its WATCH / ENFORCE switch. |
| KPI strip | IN WATCH, ENFORCING, WOULD REFUSE, WOULD CHANGE, COULD NOT MEASURE and LABELLED, over 1, 7 or 30 days. |
| Sample events | Recent requests the control would have refused or changed, with the reason, BEFORE and AFTER text where prompt capture is on (sensitive values appear as masked tokens), RIGHT / WRONG buttons and OPEN TRACE. |
| Passkey check | "Needed for every change, in both directions." Three steps, CHALLENGE, PASSKEY, SIGNED, then the signed event and its version. |
| Still enforced in watch | The protections that hold in every mode, listed on the page. |
| Yesterday | A daily digest for each control, in the console. |
A check that could not run, because its control was down or timed out, is forwarded in watch and recorded as could not measure, never as clean.
What holds in every mode
Watch is for evidence. These protections stay on whatever mode a control is in.
- People keep every restriction. Console logins and personal tokens are judged exactly as before.
- Revoked or expired keys and disabled agents are still refused, and so is any reach into another tenant.
- Quarantine, the kill switch, the emergency spend cap and any restriction an admin placed on an agent by hand still hold. The gateway's own guards, such as overload and request-size limits, always run.
- The Audit Ledger is always on: every decision is signed and appended in watch and in enforce.
- Credential Broker in watch: an agent's own provider key passes through untouched and the record says what enforce would have done. A credential the agent was never granted is never handed out.
- Cost and Budget in watch: a budget refusal is recorded as a would-refuse, and cost and tokens are metered exactly as in enforce.
Where watch first fits
Any team that has to show its work before it changes a control.
- Goals: AI visibility, data loss prevention, budget and token spend, audit and accountability, compliance evidence and agent identity and ownership. Unregistered or ownerless agents are recorded in watch, then refused once Identity Governance enforces.
- Industries: financial services, where New York's financial regulator points firms to inventorying every system that uses AI and monitoring for unusual queries;³ insurance, healthcare, public sector, retail and all industries.
- Use cases: securing AI coding agents, rolling out copilots and chat safely, controlling agent tool access and AI in regulated and air-gapped environments.
- What you enforce next: sensitive data redaction, model control, AI spend caps, agent reach control, identity tethering and one-click audit evidence. See all capabilities.
How a team starts
Watch first. Nothing changes for your agents until you decide.
- Watch. Route your agents through Argent. Every control starts in watch, and Rollout fills with what each one would have done.
- Judge. Read seven days of your own traffic, the page's default range, and mark the samples right or wrong.
- Enforce, one control at a time. Each switch needs a fresh passkey check and leaves a signed record, and you can switch back the same way. Choose where Argent runs from the deployment options, and book a demo below to see it on your own use cases.