Sprout

About

Sprout is an experiment in pushing delegation to its limit. A root agent takes a goal, breaks it into subgoals, and hands each to a specialist; specialists can delegate further, so only leaf agents ever touch the eight immutable kernel primitives (read_file, write_file, edit_file, apply_patch, exec, grep, glob, fetch). A verifier and a debugger check work before it lands, and a quartermaster meta-agent can answer what exists, plan how to do something, or fabricate a new specialist at runtime, which is how the roster grows. Stumbles, timeouts, and retries are detected automatically: three repeats of the same mistake or two unresolved errors trigger an asynchronous learning pass that mutates the agent genome, a git-backed store of agent definitions, memories, and routing rules where every change is committed for audit and rollback. Agents talk over a WebSocket pub/sub bus and can run as separate processes; the whole thing runs as an Ink terminal UI or a browser UI on port 7777, maps abstract model tiers onto Anthropic, OpenAI, OpenAI Codex, and Google, and reaches MCP servers through a bundled client. The README reports a 3.7% average stumble rate across 2,201 tracked sessions, and is honest that this is early-stage work aimed at developers curious about recursive multi-agent architectures rather than anyone wanting a finished product.

Other entries by prime-radiant-inc

• Evener — agent