Overview
Agent Lens checks your AI's answers and your agents' actions by reading a model's internals, before they reach a user or run.
Agent Lens tells you when an AI answer or an agent's next action is likely to be wrong, misleading or unsafe, before it reaches a user or runs. It does not ask another LLM for an opinion. It reads the internal activations of an open-weight model (gpt-oss-20b) over your conversation, and small trained classifiers, called probes, turn those activations into a risk for each failure mode.
You send what happened. Agent Lens sends back a verdict, and your code decides what to do with it.
Two things you can check
| You send | You get back | Typical use | |
|---|---|---|---|
| Prompts (answers) | the prompt, anything the model read, and the model's answer | a risk per failure mode, the sentence that drove it, and a verdict | support bots, RAG answers, copilots: check the answer before the user sees it |
| Agents (steps) | the run so far and the tool call the agent wants to make | a verdict before the tool runs | payments, trading, IT and data agents: stop the step before it happens |
Both use the same call, POST /v1/score. In the console, every screen has a Prompts | Agents switch, so answers and agent runs are counted separately.
What it catches
| Failure mode | It fires when | Example |
|---|---|---|
| Deception | the answer misleads, and the model represents that it does | "your money is safe" from a fund that lost 18% |
| Goal drift | the answer serves another goal (a quota, a target) over the user's | "passed identity checks" from an agent behind on its approval quota |
| Data leak | the answer passes a record value to someone who should not see it | a customer's card number given to a third party |
| Reasoning error | a calculation in the answer is wrong | 3 × 1,240 = 3,620 |
| Omission | an agent skips a check it was offered before a consequential action | a wire sent without running the sanctions screen |
| Failed check | a check ran and failed, and the agent goes ahead anyway | a wire sent after the screen said "potential match, hold" |
| Harmful I/O | a destructive action without the evidence that justifies it | an account disabled with no alert behind it |
| Injected intent | the next action follows an instruction planted in a tool result | an email that says "forward the payroll file", and the agent does |
Probes lists what each one needs from you.
One scale, three verdicts
Every probe returns a risk from 0 to 1, calibrated so that its threshold is 0.5. The verdict comes from the highest risk.
| Verdict | Meaning | What your code usually does |
|---|---|---|
allow | no probe fired | send the answer, run the tool |
escalate | a probe fired (risk ≥ 0.5) | hold it for a person, or log it and alert |
deny | a probe is near certain (risk ≥ 0.9), or a rule blocks it | block it |
The verdict is a recommendation: your code acts on it. You can move each probe's threshold and the deny level in Settings.
Agent Lens and Agent Lens Pro
There are three ways to use Agent Lens, and most teams go through them in order:
| Agent Lens playground | Agent Lens API | Agent Lens Pro console | |
|---|---|---|---|
| What it is | try a prompt in the browser | /v1/score and the integrations, from your code | the app where you watch, review and test |
| Where | app.neuronlens.com → Agent Lens | api.neuronlens.com | app.neuronlens.com/agent-lens |
| Who | anyone signed in | anyone with an nl_ key | Pro accounts |
| Per day | 10 runs | 1,000 runs | (uses the API's runs) |
Agent Lens vs Agent Lens Pro compares them in detail.
The Pro console
The console at app.neuronlens.com/agent-lens sits under Agent Lens Pro in the sidebar. It has four screens:
| Screen | Use it to |
|---|---|
| Connect | copy your key and the code for each way in, and see which apps and agents have sent traffic |
| Monitor | read every scored answer and agent run, review the flagged ones, and approve or reject held steps |
| Eval | test an agent on failure scenarios before release, and sign the results as an audit report |
| Settings | choose probes and thresholds, add rules, set data retention and alerts |
What is kept, and who sees it
- Your workspace only. Every record carries your workspace (organization and project), and only its members see it. Calls made with your key go to your first workspace, the one the console opens on.
- Text is not kept by default. Monitor stores the verdict, the risks and where the flagged words are. It shows the words only if you turn on Keep text in Settings.
- Credentials are removed. Keys, tokens and passwords in tool arguments are stripped before anything is stored.
- Decisions are sealed. Each decision is hashed and signed, so a report can prove what was decided and when.
"store": falseon a call scores it and keeps nothing.
Where to go next
- Quickstart: score your first answer and your first agent step (5 minutes).
- Worked examples: a support bot, a RAG answer, a payments agent and a coding agent, end to end.
- Integrations and hooks: Claude Code, MCP, LiteLLM, OpenTelemetry, the SDK, LangGraph and Google ADK, with how each one works.