I just published WAVES, the method I have been using to run AI agents in parallel without letting them run wild. An agent will always tell you the job is done, and WAVES treats that as a claim and checks the evidence behind it before it counts. You fan the work out to a team of agents, each one owns a slice, and the verifier is what decides when the work is done. It runs in Cursor, Codex, Claude, and Droid, and ships as a skill you can install today. The full breakdown is on the blog, with diagrams, the handoff format, and where it fits next to loops and parallel orchestration: https://www.rayfernando.ai/waves-workers-aggregate-verify-extend Skill: https://github.com/RayFernando1337/rayfernando-skills/tree/main#waves--fan-out-to-parallel-agents Give it a run on something real and tell me how it goes.

Here is part 2. I updated both the Cursor and the Codex versions, so grab whichever you run. Quick recap if you missed v1. Parallel Orchestrate turns one big task into a team of AI agents working at the same time; the lead agent splits the goal, fans out a bunch of workers, and stitches their reports into one result. v1 made you fast. Then I got curious about how the smartest model I had access to (Fable) actually pulled this off. The secret was verification. Fable spent most of its effort checking its own work so I updated the skill and it is now in v2. What's new - Verification baked in. A worker reporting "success" is now treated as a claim, not proof; the lead agent checks the evidence behind every handoff before trusting it. - The "verifier is the moat" idea. We can't train these models ourselves, but we can decide how much checking to buy at inference time, and that checking is where the real quality comes from. - Catch problems before you spawn. The skill now verifies coverage up front (counts, ranges) so no slice quietly gets missed. This alone caught a real bug in my run. - Multi-wave, not one burst. Real work arrives in waves, so the skill plans for follow-up rounds instead of pretending one shot covers everything. - A dedicated verifier agent. For high-stakes claims you can send a separate agent to fact-check, and it never sees the first agent's reasoning, so it cannot inherit the same mistake. - Verify the final thing, not just the parts. Re-run it, re-read what you wrote, confirm it actually works. Why this matters With one agent, a confident wrong answer is annoying. With a team of them, those small errors stack and quietly poison your final result. Speed without verification is just faster mistakes. v2 keeps the speed and adds the trust, which is the part that actually lets you ship what these agents produce. This is the whole game right now. You do not need to train a model or wait for the next one; you need to get good at asking, decomposing, and verifying. The power is already yours.

@Ray Fernando Hey Ray, see the attached file. This is a markdown which extracts these six types of dynamic workflows from the Anthropic official page. I fed this in as a reference doc for the skill which it must read, and I found it to be extremely helpful because it helps the main agent determine what is the best type of workflow/structure to fan out or tournaments or adversarial reviews. It just made a materially good difference. The key is that I've said that it must read this EOF (end of file) before deciding which subagents to use and structure, and to force an EOF rather than to make it deterministic. I just said I need just a read receipt, ie saying I've read this file EOF, and then continue.