Helyn Implement for coding, I work on design and specification, agentic protocol handles coding, testing, automated testing, and reviews (multiple models), process works extremely well. Enables fast iteration cycles since most issues are diagnosed in design and specification and we don't start coding until the specification is unambiguous.
From it's agents.md - Helyn Implement 3.1 defines a seven-phase software implementation protocol (project setup; research & design; design philosophy with a Subtractive Genesis Gate; performance analysis; test planning; implementation with a bug-hunter panel; security review; performance verification; QA; release readiness; and periodic project sweeps). Each phase runs an internal creator↔reviewer "quality loop" governed by the Cockpit Landing Model, where a phase is either CLEAR (GREEN) or DISCREPANCY (RED) — never a numeric score — and reviewers must be fresh-eyed and independently verify claims rather than trust specs, reports, or coverage numbers. Supporting discipline includes self-review on every fix, named (not numbered) items, explicit self-documenting signatures, and pruning artificial machinery when the substrate can enforce an invariant; a multi-reviewer panel and escalation protocol handle high-stakes or non-converging reviews, with cross-phase issue handling, findings disposition, and severity levels as scaffolding.
See https://helyn.com/follow-the-money for an example of how such a protocol works.
Currently developing a custom agent orchestration system to resolve difficult STEM problems with provenance, the type of problems that frontier models consistently get wrong.