Nested loops

Different mistakes belong on different clocks.

If every error waits for a monthly review, the system will spend the month being wrong. If every twitch can rewrite the constitution, the system will have no constitution by Thursday.

So the experiment uses nested loops. Each loop has a job, a write budget, and a stop.

Reflex

Reflex is the in-run loop. An agent is in the middle of a task. It can correct a typo, retry a tool, or stop because the next action hits a gate.

Reflex may not change identity, price, ICP, or the constitution. Reflex may not "learn" by writing durable memory. If the run produced evidence, that evidence waits for the daily loop.

This is the commit-gate idea from the agentic-evolution paper, applied at the smallest scale: the run may propose; the record commits later.

Daily

Daily is the working review. Cell owners post what ran, what failed, what they want to write into memory, and what they want to spend tomorrow.

Evaluators accept or reject those writes. Immune scans for identity drift and fake commitment. Allocators do not rebalance the company on this loop unless a cell is on fire.

A daily note that cannot name an artifact is a status mood. Moods do not become memory.

Weekly

Weekly is selection. Which cells showed commitment? Which cells burned budget and produced only drafts? Which specialists were worth the tokens?

The allocator proposes attention and budget shifts. The founder reviews irreversible items. A cell can be paused here without a theatrical postmortem.

Weekly is also where task-order and variance get a look. If the same "improvement" only works when the work arrives in a friendly sequence, that is a fragility signal, not a win.

Monthly

Monthly is constitution and kill.

Did the market still match the rules we wrote down? Did a role quietly become the founder? Did an improvement loop make the work worse? Is it time to stop a cell, stop a role, or stop the experiment?

Monthly is allowed to change the constitution. Daily is not. That sounds obvious until a useful agent asks for a "small exception" seventeen days in a row.

What a loop produces

Each loop should leave an artifact another person can read:

• Reflex: the run record, including the gate that fired.
• Daily: accepted and rejected writes.
• Weekly: a selection note and a budget proposal.
• Monthly: a constitution diff, or an explicit "no change."

If the loop produced only a meeting, it did not happen.

Current state

These cadences are the design. They are not a claim that the company is already running them with customers. Hardening and the identity canary come first. The journey log is where a later dispatch can say a loop actually ran.

Read what is not being built yet before assuming CRM, billing, or a first offer sit on these clocks.