Agent Observability
An autonomous agent you cannot inspect is a liability. Every action a Xiilio agent takes is recorded in full — what it did, why it did it, what it touched, what it cost and who signed it off.
Nine signals, captured whether the action succeeded, waited, was blocked or failed. Each one exists because a specific failure is invisible without it.
Every action an agent takes, timestamped to the second, with the agent, the customer workspace, the trigger that started it and the run it belongs to.
Without a complete log there is no way to answer 'what happened on the 14th' six months later.
What the agent was given to work with, and every system and dataset it read during the run.
A decision can only be audited if you know what the agent could see when it made it.
One plain sentence stating the decision, and the checkable facts it rested on. Never raw model chain-of-thought.
A score with no explanation cannot be challenged, corrected or defended to a regulator.
Each system the agent touched, what it asked for, what came back, how long it took, and what the agent then did or was stopped from doing.
Most agent failures are integration failures. The tool call is where you see them.
Every guardrail evaluated on the run — consent, suppression, spend caps, scope, data residency — with the result of each.
Showing the check ran and passed is the difference between compliance and a claim of compliance.
A confidence figure on every decision, the risk class of the action, and the threshold that decided whether it ran, waited or escalated.
Confidence is what lets you automate the routine 80% and route the rest to a human.
The rule that required approval, who decided, what they decided, when, and any note they left.
This is the accountability record. Named person, timestamp, decision.
What the action produced — a reply, a booked meeting, a CRM update, a budget shift, nothing at all — and the error if it failed.
Activity is not achievement. Outcome is the only measure worth reviewing.
Model spend for each individual action, the token counts in and out, and the exact model version running when the decision was made.
Autonomous systems get expensive quietly, and when behaviour changes the model version is the first thing to check.
Every action, filtered how you need it, with four dashboards reading the same records. Open any action to see the decision, the evidence behind it, every system it touched, the policy checks it passed and what it cost — then replay the whole run step by step. The activity below is an illustrative demonstration, not a client's record.
Actions taken
14
Needing a person
7
Model cost
£7.33
Tokens used
803,500
Showing 14 of 14 recorded actions.
Why did the agent do this?
Ranked 31 of 136 new accounts above the fit threshold and queued them for review; no account was contacted.
Evidence
Decision summary and evidence only. Raw model chain-of-thought is not captured or shown.
Input
148 accounts added to the source list overnight, plus the ICP definition and 18 months of CRM history.
Trigger
Scheduled — daily account refresh, 06:00
Data sources accessed
Action taken
Created 31 ranked accounts in the CRM with the fit reasoning attached to each.
Policy checks
Tools used
Outcome
31 accounts queued for review. No contact attempted — outreach requires a separate approval.
Model
Scoring model
Model version
v4.2.1
Tokens in
214,800
Tokens out
18,400
Cost
£1.94
The controls that decide what an agent may do unattended are on the Responsible AI Centre. Live approvals and system health sit in the AI Operations Centre.
Bring your risk, audit or procurement team. We will walk through the activity record action by action and answer what happens when an agent gets it wrong.
We use a single first-party cookie to remember your currency and consent choice. Page-view analytics are sent to our own servers — no third-party trackers, no profiling. See our Privacy Policy.