Bartosz Kotrys

Founding Engineer at @QuesmaOrg

[Jacek Migdal @jakozaur]

Things I shipped last weekend:

1. Offsite @QuesmaOrg presentations and photos, hosted on our own site behind SSO. Live the day after the event.
2. Animated visualizations for FlexDrill, because the founder asked.
🧵 1/4

[Bartosz Kotrys @bartkotrys]

AI token efficiency is becoming its own engineering discipline.

Not just prompt trimming, and not just choosing a cheaper model.

It is about understanding the full system:

• which model handles which task
• how much context is actually useful
• when caching helps
• where agents repeat work
• how memory changes cost
• when verification is worth the extra tokens
• how routing affects quality and spend

That is what we are trying to capture with Awesome AI Tokenomics.

The goal is to build a practical map of how AI systems spend tokens, where they waste them, and which architectural choices actually improve the cost-to-quality ratio.

It now covers tools, research, practices, claims, and reusable setups across routing, caching, context engineering, memory, observability, inference, and multi-agent systems.

Still early, and I would genuinely like this to become a community-maintained reference rather than a static collection.

Contributions, corrections, and missing perspectives are very welcome:

https://t.co/IBm7YLIWRA

[Bartosz Kotrys @bartkotrys]

@jakozaur @QuesmaOrg @pmigdal Well done @pmigdal! Big thing 🤝

[Jacek Migdal @jakozaur]

Yesterday, Frontier-Bench landed 0.1 with tasks from @pmigdal and @QuesmaOrg . Today, Opus 5 was released, and it was put in the spotlight.

[Jacek Migdal @jakozaur]

Retro today during @QuesmaOrg offsite. Time for token economy.

[Bartosz Kotrys @bartkotrys]

5/5

The hardest bug was in my own policy.

The pipeline rejected a legitimate 56k-star project because one claim was unverified.

The fix:

Reject the claim, not the project.

Agents execute your rules. They will not tell you that your rule is the bug.

Full architecture, failures, and lessons from our work at @QuesmaOrg:

https://t.co/rNsz0tDpDh

[Bartosz Kotrys @bartkotrys]

1/5

I burned my entire Claude Max 5x limit in 30 minutes researching how to save tokens.

111 agents launched.
123 claims queued.
25 verified.
No final report.

At @QuesmaOrg, I rebuilt the pipeline. It now runs roughly 10x longer across subscriptions we already pay for.
https://t.co/rNsz0tDpDh

[Bartosz Kotrys @bartkotrys]

@QuesmaOrg 4/5

I also moved deep research to the end.

First run:

111 agents
30 minutes
No final report
Later run:
61 agents
22 minutes
Only claims that had already survived verification

Same tool, much smaller and better-defined job.

[Jacek Migdal @jakozaur]

Bartosz Kotrys @bartkotrys just joined the @QuesmaOrg founding team. The path there was unusual. 1/4 🧵