Filip M

Co-founder @ Scott is the best manager of all time

[Filip M @Wanawrin]

@halluton This literally happened to me today

[Dan Kornas @DanKornas]

Agents shouldn’t make the same mistake every session.

Agentic Context Engine (ACE) is an open-source Python engine for adding a persistent learning loop to AI agents.

It helps agents reuse lessons from feedback and traces by maintaining a Skillbook: a collection of strategies that gets updated as tasks run, fail, and improve.

Key features:

• Persistent Skillbook – stores strategies that evolve across tasks instead of disappearing after one session
• Reflection loop – a Reflector analyzes execution traces while a SkillManager adds, refines, and removes strategies
• Simple quick start – install with uv, run ace setup, then use ACELiteLLM to ask, correct, and inspect learned strategies
• Trace-based learning – extract strategies from existing agent logs without re-running the original tasks
• Runner integrations – includes paths for LiteLLM, browser-use, LangChain, Claude Code, and MCP workflows

It’s open-source under the Apache License 2.0.

Link in the reply 👇

[Tony @halluton]

POV: Having to deal with slopus again, now that fable is gone

[Tom Dörr @tom_doerr]

Self-improving AI agents using persistent memory

https://t.co/AwgGzOMMzd

[Tony @halluton]

LangSmith Engine looks slick

Kayba does the same thing, but framework-agnostic and open source ⬇️

[Filip M @Wanawrin]

@JakeRoggenbuck @ycombinator @garrytan A quality build + well told story in one day, very impressive!

[Filip M @Wanawrin]

@geekbb @omarsar0 check out this autoharness!

[Geek @geekbb]

Agent harness 自动化优化工具，接管了 Agent harness 优化的脏活，你给一个基准测试命令和目标仓库，它就自动生成提案、跑评测、记结果、留好的，弃差的，自动改进 agent 的 prompt、配置和源码。

https://t.co/2qhYImGjuP

[astaxie @astaxie]

终于有工具接管 Agent harness 优化的脏活了。

你只需要给它两样东西：
• 一个基准测试命令
• 一个目标仓库

剩下的它全包——自动生成提案、跑评测、记结果、留好的、弃差的，连 prompt、配置和源码都帮你改好。

开源地址：https://t.co/gCRvFokxT3

[Filip M @Wanawrin]

@halluton It's increasingly hard for humans to pick up signal in AI written text, must be even worse for training data samples.

[Filip M @Wanawrin]

@halluton Great read

[Tony @halluton]

https://t.co/ke53KtIhGZ

[Filip M @Wanawrin]

Anthropic just fumbled the ball big time.

Having the chance of capitalizing on the influx of new users post positive public perception on their stance towards defense > it disgusted all existing and new users instead by making their models rate limited and dumb.
GG

[Nikos @kalio_late]

60 builders. 0 fees. 0 equity. 2 buildings. 1 bet.

Founders. Investors. Researchers. Operators.

@thejfloor is launching. Zurich builds.
April 16th, Zurich.

Limited spots.
Link in the comments.

[Filip M @Wanawrin]

@ta_eis_eauton That's deep man

[Filip M @Wanawrin]

Making agents do non-trivial tasks well is still such a difficult process.

[Filip M @Wanawrin]

@halluton @ta_eis_eauton There will be an agent automating this

[Filip M @Wanawrin]

@CreateWithBloom Love it, more people should know about this.

[Bloom (YC X25) @CreateWithBloom]

Apple is banning vibecoding apps.

A new generation of personal software is emerging, and Apple is blocking it from reaching iPhone users.

Everyone can build apps now. Just not on iPhone.

[Filip M @Wanawrin]

@joni_vrbt I get shredded every time I make a post on reddit and paraphrase it with AI. The arguments in the comments bring nice traffic from people complaining about it though.