Teaching Your Agent Not to Repeat a Bad Run
A two-tier scoring system: a free check on every run, and a model-written review only when you've already flagged one as wrong — an honest, cost-aware version of 'agents that learn from mistakes.'
"Agents that learn from their own mistakes" is one of the most-repeated promises in AI agent marketing right now, and most of it is vaguer than it sounds. Piskes's version is concrete enough to actually explain, and honest enough to admit what it costs.
The pattern, and the honest version of it
The general shape circulating in 2026 self-improving-agent writing is: score the run, diagnose what went wrong, write a lesson, and read that lesson back the next time a similar situation comes up. One frequently cited result shows accuracy climbing from 58.3% to 78.2% across iterations of exactly this loop. It's a real pattern with real evidence behind it — but running a model call to diagnose and write a lesson after every single run would be expensive at any meaningful volume, and most descriptions of the pattern gloss right past that cost.
Two tiers, not one
Piskes splits scoring into two tiers specifically to deal with that cost honestly.
The first tier is free and automatic: every finished run gets a deterministic score, computed from signals the system already has on hand — details already captured about the run itself, the individual steps it took, and any approvals it needed along the way. No model call, no added cost, and it happens for every run without exception.
The second tier is the expensive one, and it's deliberately rare. A full, model-generated review — the kind that actually diagnoses why something went wrong and writes a lesson from it — only runs later, and only under two conditions: a person has already downvoted the run, and the free tier hadn't already caught the problem on its own. That second condition matters as much as the first — if the cheap, automatic check already flagged the issue, there's no reason to spend a model call confirming what's already known.
Why the cost structure is the actual point
It would be easy to describe this feature the same way most "self-improving agent" marketing does — as an unconditional promise that the system learns from every mistake. That's not what's actually built, and the honest version is more interesting anyway: a free check runs constantly and catches most of what's catchable for nothing, and the expensive, genuinely diagnostic pass only fires for the cases that actually warrant it — a run a person cared enough about to flag, that the cheap tier missed.
That's a more credible design than "learns from everything, always" — because it's the version that actually survives contact with a real budget. An agent platform that claimed to run a full diagnostic review on every completed run, at scale, would either be quietly not doing that or quietly burning an enormous amount of money doing it. Piskes's version tells you exactly which of those two things is actually happening, and it's the cheaper, more sustainable one by default — with the expensive path reserved for when a human has already said something went wrong.
What this means in practice
If a run goes badly and you leave a thumbs-down, that's not just feedback disappearing into a queue — it's the specific trigger that can unlock a deeper, model-written diagnosis of what happened, if the free tier hadn't already caught it. Flagging bad runs isn't just venting. It's doing the one thing that actually escalates a run into real scrutiny.