Golem 1.6 is coming out, and the release represents a major advancement for building custom agents, harnesses, and agentic workflows on Golem.
I'll be talking about and showing many features of 1.6 throughout the coming weeks, but more instructive for now is simply the one unifying theme that ties together all of 1.6.
We all know that the most complex type of agent you can build is a coding agent. Coding agents build and run software, and therefore, in theory they can do anything that a software engineer can do--which is pretty much anything.
However, I have argued that all agents should be coding agents, because coding gives agents the ability to make custom software to solve problems they are tasked with--even ordinary business problems like, "Show me top performing accounts in a bar chart and send the top performer a congratulations email."
If all agents must be coding agents, then ultimately, then I reasoned that Golem should be sufficiently powerful to support this use case directly. As a result, with feedback from early users of Golem, I shaped the 1.6 feature set around developing a coding agent that runs entirely in Golem--durably, with a precise permission model, and the ability to add ironclad guardrails to increase reliability.
The results you will see for yourself soon enough, and I believe Golem will have attained the milestone where Golem beats LangChain and friends in any head-to-head comparison.
As proof, we will be shipping a demo coding agent that writes TypeScript, and which runs entirely in Golem, benefiting from Golem's durability and permission model.
Unlike all the other coding agents in this category, this coding agent will run entirely in WASM, not a VM, and feature millisecond startup time, consume MB instead of GB, and attain economies impossible with VM technology.
We're about to enter an age where even the most powerful type of agent--a general-purpose coding agent--runs entirely in its own sandbox, durably, starting in milliseconds, and consuming hardly any memory, enabling heretofore unimagined use cases for general-purpose agents.
Stay tuned for more.