posted an update —

HELM just crossed a major threshold

Over the last two weeks, I’ve pushed HELM from a chat/session-centered prototype toward a persistent governed state model.

What that means in practice:

  • long-running hosted agent sessions
  • durable state outside the model session
  • native subagents for independent review
  • restart and recovery without losing mission context
  • read → reason → write cycles against governed shared state
  • evidence, receipts, and Watch Floor state preserved across runs

I also pushed HELM hard in a completely different domain: genealogy.

That turned out to be one of the toughest stress tests I could have picked.

Genealogy forces the system to handle:

  • messy identity resolution
  • conflicting evidence
  • relationship uncertainty
  • source provenance
  • cluster analysis
  • long-lived investigations
  • human approval gates
  • persistent research state

And it worked well enough that it changed how I think about HELM.

At the same time, I started experimenting with HELM as a coordination layer between different AI assistants.

One of those experiments involved “Brenda,” a separate assistant being used to support HR work. The goal was not to make one assistant impersonate another, but to test whether HELM could create a governed handoff structure through shared artifacts, structured interviews, and persistent state.

That matters because I do not think the future is one giant AI doing everything.

I think it is multiple specialized assistants working against the same governed reality, each with clear roles, permissions, evidence, and human oversight.

The core idea is getting clearer:

Humans stay in authority.
Agents do the work.
State persists outside both.

Military Slices was the first proof that HELM could govern a user workflow.

Genealogy is proving it can govern a long-running investigation.

The Brenda experiment is starting to show how it could govern coordination between specialized assistants.

The next phase is proving HELM can do all three at once.

Still building. Still testing. Still breaking things on purpose.

But the architecture is getting real.