Members-Only
Recent Talks & Demos are for members only
You must be an AI Tinkerers active member to view these talks and demos.
Recursive Harness Builder
Explore recursive agent harnesses for complex, long-running tasks. This talk details a system with unlimited subagent depth, forked agents, and internal communication, enabling robust, low-supervision workflows.
I’m building an agent harness that facilitates recursive workflows for performing complex long-running tasks. The harness itself is a minimal layer on top of an agent sdk, adding to it (1) unlimited depth of subagents, (2) the ability to launch forked subagents, and (3) internal inter-agent communication protocol. These three things create a “canvas” on which, any imaginably complex workflow can be represented with markdown files referencing other markdown files.
The PoC of that is my specific set of markdown files which:
(1) tell the agent to recursively split its given task into a few subtasks (with a base condition when the task is trivially simple)
(2) run execution-verification loop at each stage until the task is done.
This is effectively a multi-dimensional ralph loop that effectively guarantees that if an agent can tell whether the task is completed, it gets completed. Not token-efficient or fast, but requires little to no supervision which makes it worth it for my use.