Plasma AI

Infrastructure for intelligence at scale.

[Plasma AI @Plasma__AI]

Introducing Fractal:

An open source tool for building hierarchical agent loops.

We built Fractal to solve our hardest problems.

[Plasma AI @Plasma__AI]

@snskritinaruka Would love to hear your feedback

[Plasma AI @Plasma__AI]

Introducing Fractal:

An open source tool for building hierarchical agent loops.

We built Fractal to solve our hardest problems.

[Plasma AI @Plasma__AI]

@ArcadianComp Great to hear, let us know how you like it!

[Andrew Turner @__acturner__]

@nicholas_diao and I just open sourced Fractal, a tool that we use internally at @Plasma__AI.

Fractal began with the question "can agent loops effectively manage other agent loops?"

[Plasma AI @Plasma__AI]

@disarto_max Incredible work as always 🫡

[Plasma AI @Plasma__AI]

Fractal is open source under Apache 2.0. We’re excited to see what you build with it.

CLI: pip install fractal
Agent skill: fractal install

Repo: https://t.co/b11Kb76gu7

[Plasma AI @Plasma__AI]

Fractals are self-organizing.

Nodes delegate work to sub-nodes, each running their own loop. Parents review and merge the output of sub-nodes before planning again, allowing work to compound across the hierarchy.

Full write-up: https://t.co/jRghTJSI2H

[Plasma AI @Plasma__AI]

@adelbucetta @karpathy Versioning and rollbacks are handled by git. The wiki is markdown, so you can branch, revert commits, or restore any page, and `wiki update` rebuilds the indexes after. This especially helps when there are multiple agents in the loop reading and writing the same context.

[Plasma AI @Plasma__AI]

We're open sourcing the implementation of @karpathy's LLM Wiki that we use internally.

What we changed: instead of the LLM maintaining the wiki, a deterministic CLI does. It keeps the index, links, and merges in shape, so the agent just writes the content.

[Plasma AI @Plasma__AI]

In practice the agent reads a compact map (each branch shows its word cost), then opens only the pages a task needs instead of the whole repo. It's plain markdown in git, so you and your agents read the same files.

Repo: https://t.co/Aw1LgYtDAk

[Plasma AI @Plasma__AI]

We build with a lot of agents at Plasma, often several on one project. As they moved from short tasks to long-running agentic loops, context management became the bottleneck. They needed a shared place to record what they learned, recover decisions, and build on each other's work.