Josh Rosen
@JoshARosen
Researching modern AI architectures: agents, factories, harnesses, context, evals, and data.
Building ThruWire. Co-founder TopCoat (acq. by @snyksec)
361 Following 2.6K Followers
LangGraph maps extremely well onto Jev.
Not because it was built for decision models specifically, but because its core abstraction is already about putting decisions into state, then routing based on them.
I wonder if over time we’ll think about LangGraph less as agent orchestration and more as just the way you build software with composable models.
Show more
I like how simple the @OpenHandsDev memory approach is for their software factory.
Agents just write to a project-level MEMORY.md file.
This is powerful in three ways:
1. It allows agent workers to leave behind state and be ephemeral.
2. It creates a basic shared-learning mechanism. If one agent learns something useful, every future agent can benefit.
3. Each agent is expected to be a good steward for the greater good, in this case the project, and not just think about its own task.
Show more
@ashwingop Love this take on memory. It’s “schema on read” more than anything else. The lens that we look at memories can change over time.
