AI AGENT THAT LEARNS
An AI agent that learns from real results
An AI agent that learns gets better at your work from its own graded results. C5 counts how often each prompt variation was graded and how often it did well, gives every new idea a fair trial, retires one only on strong evidence, and records the odds of every choice it makes. Skills are suggested to you, never assigned behind your back.
- Counts results, never averages
- Skills suggested, not assigned
- Every choice recorded
It counts, it doesn't guess
Every time a variation of a prompt is used and its work is graded, C5 records two plain numbers: how many times it was graded, and how many of those it did well. Decisions are made from those counts, never from a running average that drifts with the last few results.
- One bad day does not sink a good prompt. Five good results out of five is weaker evidence than ninety out of a hundred, and C5 treats it that way.
- Two kinds of evidence, kept apart. The grade a reviewer gives the work is counted separately from whether that review was right, so a reviewer that let a problem through takes the blame, not the work.
- Real results only. What C5 learns from is graded work and your corrections — not a guess at what you might like.
New ideas get a fair trial
A new variation starts as unmeasured, not as zero. C5 gives unmeasured ideas a fair share of trials until there is enough evidence either way, then shifts toward the ones that win. It never retires something that simply has not been tried yet.
| Decision | What it takes |
|---|---|
| Keep trying it | Not enough graded results yet either way |
| Replace its parent | Clearly better over at least ten graded trials each |
| Retire it | High confidence that it wins less than 15% of the time |
Suggestions, not silent changes
C5 tracks which skills help which kinds of agent, taking into account the other skills used alongside them, so a skill is not credited for a companion's work. When one is clearly helping a role, C5 suggests adding it: you get a notification and decide. It does not add skills to roles by itself, because those assignments are yours to make.
Specialists with a stronger record are listed first when the planner picks who does a step, but nobody is left off and the planner still chooses. A small share of tasks keeps the old order on purpose, so C5 can tell whether listing by trust actually helps.
Every choice is recorded
Whenever C5 picks between options — which prompt variation, which helper service — it records what it picked, the odds it gave each option at that moment, and what actually ran. That record lets C5, and you, check later whether a choice was a good one, without guessing what it was thinking at the time.
Learning is one half of a team that checks its own work: see how multi-agent orchestration plans, staffs and reviews a task, and the guardrails that decide what an agent may do alone.
Stress-tested on purpose
Grading the good days is not enough. In Labs › Factorial, C5 quietly re-runs a small share of steps with deliberate obstacles — time pressure, a nudge to doubt itself, a misleading hint, missing context, or the task reworded — in a balanced rotation.
The Main effects table then shows how much each obstacle costs in quality, with a range. A wide range just means there are not enough runs yet. If you want an agent that never changes how it works, this is the wrong fit: C5 is built to keep adjusting, inside the rules above.
Frequently asked questions
Does C5 change how it works without telling me?
Not where the choice is yours. Skills are suggested, never assigned to a role behind your back; fleet trials arrive switched off until you turn them on. Over time, C5 shifts toward prompt variations that win on graded results, and records every choice it makes.
Can an AI agent that learns get worse?
C5 guards against it. Decisions come from counts of graded results, not a drifting average, so one bad day does not sink a proven prompt, and a newer variation replaces its parent only when it is clearly better over at least ten graded trials each.
What does Mothership share from my computer?
Only instructions C5's own learning loop has proven, scrubbed of secrets first. Never your prompts, files, conversations or results. Mothership is optional, and C5 learns from your own use without it.
What if a fleet improvement does worse for me?
It is switched back off and what it replaced comes back, because your own results decide. Proven fleet improvements switch on by default; you can turn automatic switching off and take them one at a time.
In the docs
Let it get better at your work
Download C5 for macOS, Windows or Linux and give it the jobs you repeat. Correct it once; the next run starts from what it learned.
