LimitifyGovernance is not a policy document. It is the runtime.
Every agent Limitify AI deploys passes through the same four structural layers before, during, and after it runs. None of this is optional per-agent configuration — it is the standard every agent is built against.
The Agent Registry
Every agent is registered with a fixed structure before it can run. No agent is unregistered, and no agent's permissions are inferred at runtime — they are declared, versioned, and auditable.
Stable identity, human-readable name, semantic version — every improvement bumps the version and appends to history, never overwrites.
A–D authority level. Fixed at registration, enforced at runtime by the permission gateway.
Which client this instance serves. The mechanism that makes one agent codebase reusable across clients.
An explicit allow-list per capability (API access, storage writes, file operations). Nothing is implicitly allowed.
Log path, run counts, success/fail counts, accuracy score, average processing time — the self-learning loop's data store.
Lifetime queries, uptime start, version history — the performance record that accumulates from day one.
Authority by Category, Not by Trust
Category C and D agents never write to production or message a user directly — this is enforced structurally, not by prompt instruction. A sophisticated scanner and a simple cron job can sit in the same category if they carry the same authority.
Full authority — every state-changing action requires explicit human approval before execution.
Act within a defined, explicit permission set. Can interact with users or execute bounded tasks.
Read and report only. No production writes, no direct user contact — output always passes through human review.
Scheduled, non-interactive. No user-facing surface — pure analysis feeding the layers above it.
The Self-Learning Loop — LRBI
Every agent implements the same four-step loop after every run. This is what turns a static script into a system that measurably improves — with a human always in the approval seat for what changes.
Every run writes input, output, decision, and processing time to a structured log — before anything else happens.
Each result is scored against a pass/fail or 1–5 standard specific to that agent's job.
Errors and unexpected results write to a separate bug log with error type, input, and whether a fallback handled it.
After a fixed run interval, the agent generates a summary report. A human reviews it and approves prompt or rule changes — nothing self-modifies unreviewed.
The Analytics Standard
Performance data is collected from an agent's first run and never deleted. Monthly snapshots accumulate into a track record: lifetime queries, accuracy trend, processing time trend, and full version history — the data that proves an agent works before we ever ask a second client to trust it.