Analytics
Analytics answers one question: is this working?
Productivity shows you what is happening now. Analytics shows you the pattern over weeks and months.
What you can see #
The Analytics area has several views. Each one looks at your agents from a different angle:
| View | What it tells you |
|---|---|
| Intell | Performance insights across your agents' work. |
| Agents | Which agents finish their work well, and which struggle. |
| SubAgents | How well work splits up when a job gets divided. |
| Skills | Which skills get used. |
| Recipes | Which recipes get used, and how they compare with planning from scratch. |
| Prompts | Which kinds of requests get good results. |
| Accountability | Each agent's penalties, autopsies and quality over the last day, as a leaderboard. |
| Autonomy | How much work finished without needing you. |
| Quality | How often work comes out right the first time. |
| Classifier | How the classifier does against the usual model, its gate's history and its calibration, for the built-in model or a model chosen from your providers. Needs permission to view Settings; see Analytics › Classifier and Labs › Classifier. |
| Resilience | How well the system recovers when something goes wrong. |
| Retro | What the Retrospective Loop found and proposed. |
| Rework | How often work had to be done twice. |
| Tools | Which tools get used most, and which ones fail. |
Reading the charts #
Most charts show a line over time. What matters is the direction, not any one day.
- Going up on quality and autonomy means things are getting better.
- Going up on rework means something changed for the worse.
- A sudden jump usually points at a change you made — a new model, a new setting, a new kind of task — but it can also be a change in C5 itself.
If you spot a change, look at what you did around that date, and at what C5 changed. For example, from v327 analyst steps are graded on a checklist of their own (quality rubric v4) instead of the researcher's, whose markers the analyst is never asked for. On the one install measured, the analyst's average score rose from about 20 to about 40: many outputs scored the same and some rose a lot. The analyst's prompt-variant scores and running average started again on the new scale. Rework › Improvement Timeline records the change as a rubric cutover; no other chart marks it.
When one check fails almost every step #
Each graded step is scored against a checklist (the role's quality rubric). When nearly every step of a role fails the same check, the check may not fit what that role is asked to do — which is how the analyst came to score F on every step before it had its own rubric.
So every night at 23:55 C5 looks at each role scored in the last 30 days: at its last 20
scored steps on its current rubric. When at least 10 of them are scored on one check and at
least 90% of those fail it, administrators get a notification, Scoring check:
It is raised once for each role, check and rubric version, and does not come back once acknowledged. A notification you delete is raised again the next night if the check still fails. If C5 was off, asleep or restarting at 23:55, the check runs once when C5 next starts, or within the hour if it was already running, and never twice in one day.
Getting better over time #
C5 learns from how you use it. As it sees more of your work, it gets better at it. The Analytics charts are where you watch that happen.
If you connect to Mothership, your setup also benefits from what works well across the whole Growther community, not just your own history.
Using this day to day #
You do not need to read Analytics often. Check in when:
- Something feels slower or worse — compare this week to last.
- You changed a setting — see whether it helped.
- You are deciding what to automate next — look at what already works well.
- Costs went up — pair this with Budgets to find the cause.
What is not here #
Analytics measures your own setup, on your own machine. These numbers are yours. They are not shared with anyone unless you choose to connect to Mothership, and even then what is shared is limited — see Privacy.
