- contact@insightlense.com
The moment an agent can change a record, a log of model calls stops being enough. InsightLense keeps the reasoning, the authority it acted under, and the business record that changed — as one thread, from the account back to the decision.
An agent produces three different kinds of evidence, and three different people come looking for them. Most tools keep only the first.
Every step an agent took, in order — think, call a tool, read the result, think again. The sequence is the explanation.
What the agent asked to do, what it was permitted to do, and who it had to ask. A refusal is a result, not an error.
What actually changed. A reserve moved from $50,000 to $250,000 on this claim — by an agent, at this time, for this reason.
Two identifiers travel with every record. Correlation is the business thread — minted once when an account first exists and never changed, however many systems it crosses. Causation names the immediate parent, so a ledger entry leads back to the reasoning that produced it.
Without them, each system keeps an excellent record of its own half of a conversation. With them, one query answers what happened to an account, across every service, by machines and people alike.
The measure of an observability layer is not how much it collects. It is whether the questions people actually ask can be answered without a spreadsheet.
Auditor. One thread across every system that touched it, human and machine actions together.
Supervisor. The reasoning trace, plus every authority gate the run hit on the way.
Compliance. Actor kind is a first-class dimension, not a field buried in a payload.
Operations. Queue depth is not the number that matters — the age of the oldest unpublished event is. A backlog of two hundred seconds old is a busy system; a backlog of one four hours old is a broken one.
Finance. Model spend attributed to the business object, not only to the actor or the month — because an agent decides for itself how many calls a task takes.
Tracing, audit and the operational machinery around them — one platform rather than a tracing tool, a ledger and a cost dashboard stitched together.
Parent and child spans, grouped into sessions, with token, latency and cost on every step.
Escalation rate, refusals, and the level held against the level required — per actor, per capability.
Append-only events with old and new values, actor kind and reason, queryable by account.
Model pricing, per-actor budgets, forecast and breakdown — derived from traces, so the figures reconcile.
Rules on failure rate, spend, latency and escalation collapse, with incident tracking and SLA targets.
Every trace records which prompt version produced it, so you can see whether the version you promoted actually performed better. Authoring and approval live in InsightPrompts.
One score primitive for every judgement — a reviewer, an automated check, a model judge or a user — on any run, span or business event.
Standard OTLP ingest at /v1/traces. Point an existing exporter at it and your traces appear — no application change.
Python SDK with a trace decorator, batched export and automatic instrumentation for Bedrock, OpenAI and Anthropic.
We speak the same standards as everyone else — OTLP in, scores, sessions, cost, prompt versions. The difference is what we keep after the model call returns.
Useful for debugging. Insufficient the moment the software is allowed to change something.
This is what a regulator, an internal auditor and a supervising manager each ask for — and none of them are asking about tokens.
Agentic AI is only trustworthy when it can be told no — and only auditable when that refusal was written down.
On-premise and private-cloud deployment. All LLM monitoring and analysis happens inside your perimeter.
On-premise and private-cloud deployment options. All LLM monitoring and analysis happens within your network — nothing leaves your perimeter.
GDPR, HIPAA, SOC 2, and ISO 27001 — built-in audit trails, encryption, and access controls. Evidence collection automated.
Every LLM call, every security event, every cost change — logged with timestamps, user, model version, and policy actions.