A braid is a complex structure or pattern formed by interlacing three or more strands of flexible material. It's the name of a new TUI (github.com/tangle-network/br…) I've built over @tangle_network's agentic libraries: agent-runtime, agent-eval, agent-sdk, and more. The TUI interlaces a variety of things together, which aptly describes why its named this way. Foundational AI companies will continue to innovate on their own harnesses, improving context management, tools, RLMs, subagents, and more. This is not an easy area to compete. Moreover, these improvements are everything we hope to leverage and stay up-to-date with. Agent Runtime (agent-runtime) The purpose of the agent-runtime is to sit on top of any existing and future harness and run it locally against your subscriptions using a cli-bridge or sandboxing infrastructure like Tangle's sandboxes (sandbox.tangle.tools/). Much of parallelizing work, research, and software development is moving towards massively scalable cloud agents. This runtime sits above any major harness and gives users an easy way to run @claudeai, @openai codex, @opencode, @earendilworks pi, @PrimeIntellect agent, @NousResearch hermes, and more in the cloud. A core part of the agent runtime is a meta-improvement loop that analyzes agent rollouts/traces using agent-eval library primitives to improve and modify an AgentProfile. An AgentProfile is a standard configuration for an agent running in a VM, specifying the harness/model/tools/skills/VM snapshot expected to execute. The runtime exposes a supervisor MCP that gives these coding harness agents the ability to spawn subagents in new sandboxes. The goal is to leverage both in-harness features like ultracode, dynamic workflows, parallelization within a single VM and enable agents to manage compute across VMs. Distributed agent managed compute will allow AI scientists to run experiments in parallel and AI SWEs to build and audit components in different hardware. Spin up sandboxes, run agents with any harness, and spin them down upon completion, trace runs, improve profiles, and repeat. Braid Braid is meant to be a versatile TUI that sits on top of harnesses, routing through a OpenAI compatible local bridge to leverage your active subscriptions or Tangle's sandboxes to leverage cloud agents and parallelism. It has features that easily let you fork and ask questions to trace analysts about your session, to meta analyze and self improve over the harnesses themselves or to easily branch into new threads, a feature I long for in other harnesses. This is actively being worked on and if this is interesting to you, please reach out! My goal is not to reinvent the wheel. This is a harness on top of harnesses, which I'm sure people are building.. Happy monday! Links - github.com/tangle-network/ag… - github.com/drewstone/cli-bri… - github.com/tangle-network/br…