Every run an agent makes writes a trace: the trigger, the facts read, the rule version, the model calls, the limits checked, the action and the acknowledgement, each with the time it took.
Five things happen during a run, and all of them are written as they happen.
A trace is written while the run proceeds. The trigger is recorded with the minute it fired. Each step that follows adds its own entry: the facts read and their systems, the rule version that evaluated them, the model calls with their route and cost.
Limits are written with their result, so a reader sees what each check returned. The action and the acknowledgement close the trace, and the outcome verdict joins it when its check window closes.
Those entries feed three views. The run view is the trace itself, opened from a decision, a line or an approval queue. The agent view sums a week; the fleet view puts every agent in one table.
Counting per 1,000 decisions is what makes the fleet view readable. An agent that makes 40 decisions a week and one that makes 900 sit on the same scale.
Eight entries, each with its own timing, written while the run proceeds.
A run view answers what a planner asks about one recommendation: what it read, which rule applied, what it checked and how long each part took. The timings matter as much as the verdicts, because a slow step is usually a source system under load.
The same view opens from the other direction. A buyer on one purchase order line sees the run that touched it, and an auditor on one approval sees the run that produced the draft. The audit trail page shows the stored record behind it.
Showcase estate. The three conditions matched 38 lines; the $25,000 value test split them into 29 drafts and 9 lines that reached D. Okafor as buyer, for his judgement, with the reason for each. The waiting step is the queue time, held separately from the agent’s own time.
The same five columns would describe every agent built from the blueprint, whichever chain it belongs to and whichever level it runs at.
| Agent and level | Decisions | Approved as drafted | Overrides per 1,000 | Model cost per decision |
|---|---|---|---|---|
| A2 PO cancellation · six weeks | 212 | 96% · 203 as written, 9 edited | 0 | zero at run time · $1,284 at design time |
| A2 Safety stock · designed | yet to be built | · | · | · |
| A1 Lead time drift · designed | yet to be built | · | · | · |
| A3 Excess stock · designed | yet to be built | · | · | · |
| A2 Shortage triage · designed | yet to be built | · | · | · |
Showcase estate. The report shows what the product would report once several roles are built: the PO cancellation agent runs in production and its row is measured; the other four rows are designed roles from the agent blueprint, yet to be built, with their figures left empty until each is built. productionblueprint
Autonomous corrections count the actions an agent completed inside its limits. Approved drafts count the actions a person released. Overrides count the recommendations a person rejected or replaced, edits are counted beside them, and exceptions count the cases an agent handed to a person: at A2, the lines above the value threshold that reach the buyer. Every one is reported per agent and per 1,000 decisions.
Read together they say whether an agent is ready for more. A high share approved as written with few overrides is the evidence a level review asks for; a rising override count is the signal to read the rule again.
Cost per decision sits beside them. The token manager meters every model call, so the figure is a measurement, and read against the value an agent produced it answers the question a finance lead brings.
McKinsey, 25 August 2026, found 40% of large organisations scaling agents and 37% reporting any effect on earnings. A report that counts decisions, corrections and cost is how a manager moves an agent from the first group to the second.
Six alerts, each with a line your team sets and an owner it reaches.
Each alert names the agent, the measure, the line it passed and the owner it went to, and closing one is a decision with a reason.
Deloitte, 24 April 2026, names real-time monitoring as one of three controls most organisations still lack, beside clear decision boundaries and audit trails of the full action chain. These six alerts are that monitoring, running on the agents themselves.
Traces leave Juno in the shape your own tooling reads. Forrester, April 2026, advises adopting the OpenTelemetry conventions for generative AI now, while they are still experimental, so that agent traces sit beside application traces in one place.
Analysts ask for logs, metrics and traces for agents, together with an agent registry. Juno’s registry is the list of every agent built from the blueprint, with its owner, level, limits and measures, in one list; the blueprint beside it is the map of what can be built next.
Showcase estate. The PO cancellation agent is the one built agent; the four roles beside it in the report are designed in the agent blueprint, yet to be built, with their figures left empty until each is built.
J. Ridley opened the agent view for the level review on Friday 30 October at 16:10. Six weeks, 212 decisions: 203 drafts approved as written, which is 96%, 9 edited, 0 rejected, 0 overridden, 66 exceptions handed to a person and 212 write-backs confirmed. The share approved as written had stood above the 95% line the team had set for a move to A3 for four consecutive weeks, weeks 3 to 6.
Two figures settled the rest. Model cost stood at $1,284 for the eight weeks, all of it at design time and zero at run time, and 0 alerts were open on the agent. J. Ridley released lines under $25,000 to A3. The fleet view beside the agent view held one measured row; it will hold one row per role as each is built.
Registered on Monday, answering on Wednesday, acting on Friday: what each day holds, who you need and what you have by the end of it.
The difference is between counting what ran and measuring how well it ran.
Automation reporting usually counts events: jobs run, messages sent, records touched. Those counts answer whether the machine worked, and leave open how often people corrected it, how much it cost and whether it is ready for more.
Juno reports the corrections beside the actions, per agent and per 1,000 decisions, with the cost of the model calls beside both, so a manager compares an agent in purchasing with an agent in logistics.
For the retention period your policy sets, the same period as the audit records they belong to. Older traces export to your own store first.
Yes. Traces export in the shape your tooling reads, and Forrester advises the OpenTelemetry conventions for generative AI as the target to aim at.
One recommendation an agent produced, whether a person approved it, edited it, rejected it or the agent completed it inside its limits. Every measure counts on that unit.
The named owner of that agent, through the route your team uses, with a copy in the manager report. The alert carries the measure, the line and the runs behind it.
Each one builds on the one before it.
Reads your rows in place.
Connect and move ↗One record per real thing.
Master and govern ↗Apps for every decision.
Operate and decide ↗Asks and acts, at your level.
Ask and act ↗Bring two systems and one decision your planners make every week. In the first week we register the data, confirm the mapping with your team, answer your questions in ChatLima and run the first agent at the level you choose.
