
When a trusted agent stops acting like itself
An agent's credentials can stay valid while its behavior changes. Permission rules check what an agent is allowed to do, not whether it is still acting like itself.
- OWASP's Top 10 for Agentic Applications 2026 lists rogue agents (ASI10), and identity and privilege abuse, where leaked credentials let agents operate far beyond their intended scope.¹
- For broker-dealers, FINRA's 2026 oversight report has a new GenAI section that covers AI agents, naming risks that include agents acting beyond the user's actual or intended scope and authority.²
- For DFS-regulated companies, the New York Department of Financial Services AI guidance points to monitoring for unusual query behaviors that might indicate an attempt to extract nonpublic information.³
Picture an agent that pulls statements and drafts case notes. One morning it starts calling a payments export it has never touched, in a burst. Every call is allowed. The question is whether anyone notices.
One baseline per agent, learned from its own traffic
Argent learns what normal looks like for each agent, notices when one stops behaving like itself, and gives a person, or the policy your team chose, the next move.
- It watches how an agent behaves, not what it says. The baseline holds counts and fixed names only, never prompts, raw network addresses or content.
- Ten behaviors per agent: tool set, requests per minute, tokens per minute, hours of activity, scopes used, source networks, source geography, client class, client signature and credential context.
- One checkpoint. Argent sits between your agents and the models and tools they use, so every request it handles also teaches it that agent's normal. It covers every agent that works through Argent.
- No setup per agent. An agent is enrolled the first time it sends a request.
How it works, step by step
- Enroll. A new agent is enrolled on its first request through Argent.
- Learn. The agent shows WARMING while Argent builds its baseline. In the console's own words: "An agent earns a baseline after 50 observations of real traffic." Then it shows CLEAR.
- Check. Argent checks each agent's behavior continuously, in the background, outside the agent's requests.
- Two checks must agree. A statistical score and a learned model each compare recent activity with the baseline. Both must cross the line your team set, or one must see an extreme change, before drift is recorded, so one noisy signal does not raise an alarm. A sudden surge stands out.
- Record and show. Drift becomes a signed record and a live console event. The agent turns DRIFTING and names up to three behaviors that moved, baseline versus now.
- First-seen. Separately, the first time an agent uses a tool, network, region or client it has never used, Argent files one signed record for it, even when its counts look normal. Each one lands in the console for a person to review.
What your team sees
Everything lives in the console's Drift view and on each agent's Behavior tab.
| Where | What it shows |
|---|---|
| Drift view | Every agent with its state: WARMING, CLEAR, WATCH, DRIFTING, HELD (manual) or HELD (auto), sorted by "State severity, then drift". |
| Drift chart | In the agent's drawer, "Drift over N checks": its drift score against two dashed lines, engage and release, a red dot where drift was confirmed and a red bar for the time it was held. |
| WHAT MOVED | Up to three behaviors that changed, such as "Tool set" or "Requests per minute", baseline versus now. |
| Behavior tab | On the agent's own page, its highest drift score per day. |
| Signed records | Every verdict, hold, release and re-baseline becomes a signed record. When a person holds, releases or re-baselines an agent, the record is signed before the change takes effect. |
| What the agent sees | When held in Enforce mode: "This agent is in Hardened Mode, so Argent SI blocked the action." |
Who decides: a person, or a policy you chose
By default Argent records and shows drift, and a person decides. Nothing is held automatically unless your team chooses it.
- Hold. From the agent's drawer a person holds it for 1 h, 4 h, 24 h or until released, with a required reason. In Enforce mode the gateway refuses the agent's next request. The agent keeps its key, so a release puts it straight back to work.
- Accept the new normal. If the change is legitimate, ACCEPT NEW NORMAL re-centers the baseline on current behavior, and RE-BASELINE relearns it from scratch. Both need a reason and are signed first.
- Release. RELEASE lifts a hold. A release by someone ranked below the person who placed the hold is refused, and that refusal is signed too. Auditors can read the Drift view but cannot act.
- Automatic holds, opt-in. Teams can choose, per tenant through the Argent API, to have Argent hold a drifting agent automatically on the first confirmed drift. It releases slowly, only after sustained normal behavior; new drift restarts the count, and a hold placed by a person is never released automatically. In Watch mode the hold is recorded and the request still goes through.
- Runaway loops. A separate fixed rule catches an agent repeating the same tool step. In Watch mode it is recorded on the Loops page; in Enforce mode the request is refused with "agent loop detected: the same tool step keeps repeating".
- Tenant walls. Each tenant sees and acts on its own agents only.
Where it fits in your program
Behavior watch is one layer of the same checkpoint that governs every agent's requests.
- Goals: See all AI in your company, Agent identity and ownership, Threat defense, Audit and accountability, Compliance evidence. All goals
- Use cases: Securing AI coding agents, Controlling agent tool access, Governing customer-facing AI assistants, AI in regulated and air-gapped environments.
- Industries: Financial services, Insurance, Healthcare, Public sector, Retail. All industries
- Works with: Watch before you enforce, Agent reach control, Identity tethering, Keyless agents, Session replay, One-click audit evidence. All capabilities
How a team starts
Watch first. Nothing changes for staff until you decide.
- Switch it on. Turn on behavior watch in your Argent deployment; agents enroll themselves on their first request.
- Watch. Argent learns each agent and records drift and first-seen findings while requests pass unchanged.
- Tune. Admins set the engage and release thresholds and the minimum observations for their tenant in the console.
- Enforce, area by area. Switch on refusals when the team is ready, and opt in to automatic holds if you want them.
- Inside your walls. Argent runs in your Kubernetes cluster, as a single-node bundle, or as an air-gapped bundle verified before import. Compare deployment options