FluxPM

Modern engineering teams are increasingly mixed — humans plus AI agents contributing code, requirements, tests, and documents. The project-management tools they use were built for humans only. The default mode for AI work is to treat agents as smart autocomplete, with no acceptance criteria, no independent verification, and no audit trail. The result is the kind of drift, sycophancy, and untraced output that would never be tolerated from a human contributor — but is somehow tolerated from an LLM because "it's just a prompt."

There was no project-management platform built around the assumption that AI agents are team members who need management — onboarding, scoped tasking, acceptance criteria, IV&V, configuration control, and audit trail — exactly like the humans they work alongside.

FluxPM treats AI agent identity as first-class. Agents have API keys, scoped project permissions, and contribution traces in the same model as human team members. The platform implements three-agent orchestration: a developer agent that executes work, a UAT agent that verifies it, and an arbitrator role that resolves disagreements between them. Formal session identity and project access controls prevent agents from operating outside their scope.

The product itself is documented to the same standards it imposes on its users: an IEEE-aligned BRD / SAD / SRS / StRS / SyRS document set, with traceability from stakeholder need to acceptance test. Crucially, the V&V suite — 446+ Gherkin scenarios spanning Selenium, REST Assured, and Playwright accessibility / visual regression — was authored by a separate AI agent that has never seen FluxPM source code. Proper black-box testing, not the comfortable echo-chamber kind.

Self-hosted, in production at fluxpm.brumbiesoft.org. Used internally to manage every Brumbiesoft client engagement and product development workstream. Independent UAT suite passing on each release; release gates are spec-traced, not subjective.