Turn a goal into finished work, on its own.
Agentryx turns a goal into finished work on its own. Give it a goal; it decides whether to run the task directly or split it into a sub-task DAG, runs each piece in its own git worktree, and synthesizes the result — replanning with the actual error in hand when a step fails. It drives the Claude or Codex CLI, and it is built one small, verifiable stage at a time.
A task either runs directly or splits into a sub-task DAG, by the same recursive rule at every level. Single-agent execution isn't a special path; it's the base case of that rule.
When a sub-task fails, its parent replans with the error and partial results in hand. If it still can't recover, the failure propagates up and a higher level replans with broader context. The nearest prior method re-splits without ever seeing the error.
The closest validated peer is Anthropic's orchestrator-workers; Agentryx is its recursive generalization. It is built one stage at a time — Stages 00–02 are done (single agent, web frontend, task decomposition); agent-to-agent delegation, automatic verification, and parallel execution come later.
The README states it plainly: "Stages 00–02 are built — the task model is implemented as of Stage 02." Stage 00 is a single agent finishing one task; Stage 02 decomposes a task into a sub-task DAG, runs it, and synthesizes the result.
Agentryx drives the Claude or Codex CLI as interchangeable adapters, and each task runs in its own git worktree under a scratch repo — so an agent's work is isolated and can be inspected after the run.
The README is explicit about scope: at these stages there is no agent-to-agent delegation, no automatic verification, no parallel execution, and no multi-tenancy — those are later stages, and the public copy doesn't claim them.
An AI company should be a small operator and a lot of leverage.