Kacper Włodarczyk

Principal Engineer at Vstorm · Pydantic AI contributor · Building AI agents that run in production

[Kacper Włodarczyk @wlodarczykacper]

1,000-row sales CSV to tables and a chart in AgenticOS. 📊

I asked for Python, refunds as negatives and monthly net revenue by region. The file check wasn't in my prompt.

Every total matches a separate script: 102,826 net.

Built at Vstorm, open source: https://t.co/B6Pl50vxkT

[Kacper Włodarczyk @wlodarczykacper]

Sick design! 😎

[Kacper Włodarczyk @wlodarczykacper]

@marcelkargul https://t.co/DEOi3w2O0T

Look at this! 🚀🚀

[Kacper Włodarczyk @wlodarczykacper]

Paper: https://t.co/ABGPVuA2Li by Jiajie Wang et al.

Restating a rule the Skill already had did little. Loaded on demand instead of preloaded, agents kept about half the gain (51%). And only about a quarter of later commits needed the added rule.

[Kacper Włodarczyk @wlodarczykacper]

Does editing a SKILL.md change what a coding agent does?

A preprint replayed rules added in real Skill revisions on 80 sandbox tasks. Four agents took the required action 23 points more often, mostly when the rule named a command or path the old Skill lacked.

[Kacper Włodarczyk @wlodarczykacper]

@cyrilXBT Topic files drift once the code moves. I'd keep entries that name a command or a path, like `make test` or `src/db/`, so the consolidate step can check them against the repo instead of merging stale notes.

[Kacper Włodarczyk @wlodarczykacper]

@ggwplabs 34s to 2.8s on a snapshot needs a second check: do the faster queries return the same rows as the old ones? With ten agents tuning in parallel, a dropped filter looks like a speedup.

[Kacper Włodarczyk @wlodarczykacper]

@madmecodes When the trim kicks in matters too. My pydantic-ai-harness PR #465 resolves compaction thresholds against the model's context window, since a fixed token count fires too early on a 1M model and too late on 128k.

[SkalskiP @skalskip92]

damn, gemini 4 argon is really good at instance segmentation.

masks including holes and disconnected parts of the same object.

↓ more segmentation results

[Kacper Włodarczyk @wlodarczykacper]

@_lewtun Scoring a refusal as a wrong answer mixes safety policy into the capability curve. Put the refusal rate next to accuracy on answered items and you'd see whether max effort is actually worse or just more cautious.

[Kacper Włodarczyk @wlodarczykacper]

@Al_Grigor Hugo's split puts review on the agentic side, and that part doesn't shrink. I'd give Claude Code or Codex bounded changes with tests and keep architecture and security-sensitive code with a person.

[Kacper Włodarczyk @wlodarczykacper]

@arankomatsuzaki Keeping the patched function is where this differs from a coding agent with tests. Does CRIS-0 re-run the patch on the earlier demos, or can a fix for one IK limit quietly break a pose that used to work?

[Kacper Włodarczyk @wlodarczykacper]

@wunderwuzzi23 Windows strips trailing dots from file names, so a blocklist that compares raw strings misses `settings.json.` while the write still lands on `settings.json`. The check has to run on the resolved path.

[Kacper Włodarczyk @wlodarczykacper]

@neil_xbt Drafts only protect new objects. Flipping an existing ad set to ACTIVE or multiplying its budget changes live spend right away, so those calls need a human approval in the client. The OAuth scope alone won't stop them.

[Qwen @Alibaba_Qwen]

Congrats to @lightseekorg on the impressive TokenSpeed results! 🚀 Great to see Qwen3.8-Flash-Next reaching 685 output tokens/s per user on Nvida GPUs. Excited to see the community unlock more of Qwen’s inference potential!

[Kacper Włodarczyk @wlodarczykacper]

@pydantic With Argon continuations stitched into one run, does usage add up across every chunk? Then `UsageLimits` would cap the whole output, not each 32K piece.

[Kacper Włodarczyk @wlodarczykacper]

@anupamrjp https://t.co/B6Pl50vxkT

[Kacper Włodarczyk @wlodarczykacper]

@lightsilver323 https://t.co/B6Pl50vxkT

[Kacper Włodarczyk @wlodarczykacper]

@anupamrjp https://t.co/B6Pl50uZvl

[🚨 AI News | TestingCatalog @testingcatalog]

DAILY AI BRIEF 🗞 — Oct 13

META 🔥:
- Meta released Muse Wonder, its first frontier model, on the Meta Model API and in Muse Code at $4/$20 per million input/output tokens.
- Muse Wonder Full, Meta's max-reasoning variant, is launching only for Muse Code subscribers and won't be offered on the API.

SPACEXSI 🔥:
- SpaceXSI introduced The X Subscription, one set of plans covering Grok, Cursor and X Premium: Plus $8, Premium $30, Super $100 and Ultra $200 per month.

GOOGLE 🔥:
- Gemini 4 Argon is set to arrive today for Google AI Ultra subscribers.
- Antigravity CLI 1.3.4 adds Gemini 4 Argon model variants in low, medium and high.
- The Gemini app is merging its Spark agent into the main chat with new "Chat" and "Task" toggles and a "Scheduled" page.
- Google is preparing Gempix 3 Pro, likely Nano Banana 3 Pro, built on Gemini 4 Argon.
- Google launched Fitbit Edge, a $179 fitness tracker with a display and GPS, along with Google Health app updates.

ANTHROPIC 🔥:
- Claude Opus 5.5 is reportedly sometimes routing to a newer Anthropic checkpoint with stronger coding and 3D outputs.

MICROSOFT 🔥:
- Microsoft announced Maia 300, its third-gen in-house AI inference chip, rolling out to production in 2027.

OPENAI 🔥:
- Codex got another global usage-limit reset on Day 8.
- OpenAI reopened applications for its Codex Ambassador program.

NVIDIA 🔥:
- NVIDIA is opening its Vera Rubin architecture to the Open Compute Project ecosystem to help scale AI factories.

AMAZON 🔥:
- Amazon Bedrock AgentCore now has consumption-based pricing for short-term memory.

* Used Grok to compose this brief, cherry-picking the news and doing some post-editing.

[Kacper Włodarczyk @wlodarczykacper]

@adisingh Congrats on shipping AgentMail Evals. Scoring model and harness as pairs across 25 combinations is the right unit, since the same model can land very differently depending on the loop around it.