BARA LABS
Agent Monitoring by BARA Labs

Know what your AI agents are doing, and whether to trust them.

BARA Labs develops solutions for the safe adoption of AI, starting with agentic AI. Our first product, Agent Monitoring, is designed to record every action an AI agent takes, score its risk using methods proven in financial crime detection, and tell security teams which agents to approve, restrict or stop.

See how it works Join the pilot programme
Agent network Illustration
Your organisationOutsideAgent Monitoringsupport-01devops-03hr-02finance-07CRMCI testsCode repoCV inboxHR databaseEmailInvoicesPaymentsgmail.comsupplier.co.uk
Drag any node
  1. 09:41:07support-01crm_lookup → customer record0.2%Allow
  2. 09:41:05devops-03run_tests0.1%Allow
  3. 09:41:03finance-07read_invoice0.3%Allow
  4. 09:41:01support-01reply_customer → gmail.com0.4%Allow

An illustration of how Agent Monitoring will present agent activity, not live data. Scores are designed to be calibrated: 88% would mean about 88 in 100 such actions are harmful.

Why now

Agents have the keys. Few organisations can see what they do with them.

Agents now take actions

They read files, send email, change records and make payments, using real credentials, often without a person in the loop.

Guidance asks for monitoring

The UK's National Cyber Security Centre advises treating agent activity like user activity and alerting on unexpected behaviour in real time.

The risk is still unmeasured

Security teams have no open, evidence-backed way to turn agent activity into a risk score they can make decisions with.

How it works

Transaction monitoring, rebuilt for AI agents.

Banks have watched every payment for decades: know the customer, learn what normal looks like, score the unusual, escalate the risky. We apply the same discipline to agents.

  1. Capture

    Record every agent action through open standards: OpenTelemetry, the Model Context Protocol, or a few lines of Python. Arguments are fingerprinted, never stored.

  2. Score

    Combine auditable rules, each agent's normal behaviour and a calibrated model into one risk score, with reasons.

  3. Decide

    Give the security lead a recommendation for each agent: approve, approve with conditions, restrict or suspend.

  4. Prove

    Measure accuracy on an open benchmark, so every claim about the system can be checked.

Read how Agent Monitoring works →

What we will measure

The targets our pilots will test.

80%+of harmful agent actions caughtAt the alert volume your team can review.
50%fewer false alarms than rule-based alertingAt the same detection rate.
≤0.05calibration errorSo a 30% risk score means about 30%.

Targets for pilots with UK organisations from 2027. Results will be published alongside the method used to measure them.

Principles

Built in the open, built in the UK.

Open by default

The engine, rules and benchmark will be released as open source under Apache-2.0, so anyone can inspect how a score is reached.

Evidence over assertion

Scores are designed as calibrated probabilities with reasons, and every accuracy claim will come with a public test.

Private by design

We need metadata, not content. Tool arguments are fingerprinted, and pilots can run entirely inside your environment.

Mapped to the guidance you follow

Rules and checks map to NCSC's agentic AI guidance, the Five Eyes risk categories and the OWASP Top 10 for agentic applications.

Pilot programme

Running AI agents? Pilot Agent Monitoring with us.

We are looking for two or three UK organisations already using AI agents to pilot Agent Monitoring in 2027. Pilots are free, run inside your own environment and use metadata only.

Email us with a line about your agents: