About

Agent Shotgun

Explore a wide problem space with many inexpensive, independent agents, then synthesize their useful findings into one clear, reviewable result.

Coming soon. This initial prototype is a Codex skill and a short field note from one 30-task run. The workflow has not been validated across repeated large runs. It does not launch agents, track a live queue, or integrate changes automatically.

For work where one focused line of investigation could miss useful alternatives, the skill helps design a genuinely diverse task list, coordinate independent work, recover without duplicating it, review evidence, and integrate only useful results.

Intended outcome

A broad task portfolio should cover meaningfully different avenues rather than repeat one fixed list of subtasks. Each assignment should be small enough to understand and have a concrete deliverable, while the overall batch can tackle a complex goal. Negative and inconclusive results are useful when they close a real avenue or show what remains unknown.

The coordinator should be able to distinguish work requested, admitted, running, completed, reviewed, and integrated. A completed agent turn alone does not show that the requested work is complete. When agents touch shared source, one integrator reconciles their changes against the latest target revision.

In this repository

• SKILL.md — the initial Codex skill and operating workflow.
• EXPERIENCE.md — observations from the first large batch, clearly separated from candidate next-run practices.
• examples/bounded-wave-fixtures.md — synthetic three-lane planning and recovery cases; no native execution or concurrency claims.
• product.json — the current portfolio product metadata format.

The skill is based on one batch and should be treated as an initial working guide. It makes no promise of unlimited concurrency, guaranteed admission, or blocker-free integration. It is operating guidance, not a launcher, framework, daemon, or testing platform.