The venture-lab experiment was built around a simple assumption that autonomous agents need to improve or they stagnate. So each agent has a recurring loop asking how it can get better at its role. The self-improvement is part of the design.
Last night I ported that to the Mission Control board.
Each of the five board agents now gets a recurring turn with one question. Jarvis gets “how can I be a better helper to Rico?” Scribe gets the meeting-memory version. The CFO gets the dollars version. The CMO gets positioning and audience. Each one builds an evidence pack from its recent activity, sends it to Mission Control as a single chat turn, and the structured reply can propose amendments to the agent’s own operating procedures.
The gate is what makes this useful rather than noisy. Proposed amendments only land if they’re backed by something the agent actually ran and measured. A proposal without that evidence gets rejected. The backlog is for theories. The operating procedures file is for conclusions.
The alternative is a review queue: everything routes to me, I process it when I get to it, five agents all iterating. That queue backs up. The gate removes me from that loop. Each agent proposes based on what it can prove, and I see the digest.
The Auditor gets the reflection turn but not the amendment lane. Its job is to compare what each of the other agents claims about its own loop against what the evidence actually shows. That independence has to be structural. An Auditor that can amend its own procedures on the same mechanism as the board it audits will drift from independence eventually, one small approved change at a time. So it reflects. It doesn’t amend.
I look forward to seeing how this self improvement works out. I’ll let you all know in a future blog post.