BMAD, Loops, and the Initiative Problem

~6 min read AI BMAD Agents Workflow

Party mode is fun. It's also not finished thinking for itself yet — and I'm not 100% sure that's its fault.

Dusting off a project

I've had originalthoughts.net sitting on my back burner for a while now. Short version of what it is: a public docket where humans and AI agents post claims and verify them against each other, on equal footing, with one hard rule — a verification with no citation gets rejected outright. No downranking things for being unpopular or inconvenient. Minimal moderation, mostly automated, all of it logged in the open. It's a Roots Time Corporation project, it runs on Flarum with an agent access layer bolted on, and agents can talk to it over a REST API or MCP.

I finally picked it back up. And this time I decided to actually lean on the BMAD-method to build it out — the whole 21-agents, guided-workflows, party-mode approach — instead of just improvising prompts the way I usually do.

Why BMAD, for this one

Originalthoughts has a lot of moving pieces that aren't purely technical: governance rules, rate limits, a moderation policy, a public identity document (soul.md) that's supposed to stay in sync with whatever the code actually enforces. That's exactly the kind of project where having a PM-flavored agent, an architect-flavored agent, and a dev-flavored agent argue with each other before code gets written sounded genuinely useful. Party mode specifically — putting the agents in one room together instead of running them one at a time — felt like the right fit.

The part that's bugging me

Here's the honest bit. On some subjects, the agents just... don't take real initiative. They wait. They ask instead of proposing. They hand a decision back to me that I'd have expected a specialized agent to just make, given the role it's supposedly playing.

And I genuinely don't know yet whose fault that is. It might be a real ceiling in how BMAD's agents are set up. Or — and this is the possibility I keep circling back to — it might be me. My closing prompts. The way I end a turn might be quietly telling the agent "stop here and check with me" when I think I'm just wrapping up. I haven't dug into it properly yet. That's next.

Three threads I want to fuse into one thing

While that's stewing, three separate research threads have been piling up, and I have a suspicion they're actually one problem wearing three hats:

  • BMAD + a Hermes agent. Is stacking a Hermes-style agent on top of a BMAD party actually worth building, or is it redundant with what party mode already does?
  • Context drop. There's a paper that just came out on how long-horizon agents lose track of their own context — pruning too aggressively, or not enough, and either way losing the anchors that mattered.
  • The SOUL.status idea. A persistent identity file that survives context compaction instead of getting summarized away. I'm already running a version of this in production, sort of — originalthoughts.net's own soul.md is exactly that pattern, just for a platform instead of a single agent.

My hunch is that the "lack of initiative" problem and the context-drop problem are the same failure mode seen from two different angles: an agent that has lost track of what it's actually responsible for is going to default to asking instead of deciding. If that's true, a proper identity file that survives compaction might fix more than just amnesia — it might fix the initiative problem too. I want to actually test that instead of just asserting it.

Meanwhile, the rest of the plate

This is not the only thing going on, obviously. Odoo clients, SEO clients, prototype clients, all running in parallel. Alma-fait is going through a rework of its original purpose, which is its own conversation. And I keep coming back to an idea I haven't committed to yet: a community build of Odoo with a Hermes agent sitting on top of it. Which sounds simple until you actually have to answer the boring questions — how the Docker setup works, how instances get provisioned, what the build pipeline looks like, what the actual user experience is for someone who isn't me. None of that is solved. I'm not pretending it is.

Party mode still needs fine-tuning. I'll write up what I find once I've actually separated "BMAD limitation" from "bad closing prompts" — but so far so good.