THE SIX LEVELS OF AN AI-NATIVE ORGANIZATION Nate Busa / CTO, AI and Automation Python 3.11+. Download all seven .py files and requirements.txt into one folder. Setup: python -m venv .venv source .venv/bin/activate # Windows PowerShell: .venv\Scripts\Activate.ps1 python -m pip install -r requirements.txt Run each file directly: python l0_theater.py python l1_personal.py python l2_team.py python l3_connected.py python l4_compounding.py python l5_initiative.py Default decisions are typed fixtures. No model calls, credentials, API fees or real enterprise API writes occur. SQLite stands in for company systems. The graph/control-flow code runs normally. These fixtures test the mechanism, not AI reasoning quality, autonomous discovery or organizational maturity. Optional live mode (explicitly enables paid model calls): export AUTONOMY_LIVE=1 export AUTONOMY_MODEL= export OPENAI_API_KEY= For PowerShell use $env:AUTONOMY_LIVE = "1" and the corresponding variables. Do not put credentials in source files or commit them. The local adapters remain. The live provider path is included but was not run for the article's verification. Expected fixture behavior: L0: announcement only, no model or customer-case action. L1: a draft based on the pasted text; no shared system access. L2: one local support queue assignment to engineering. L3: displays a proposed internal account note; waits for y/N. Refusal writes no business action. SQLite checkpoint + thread ID let the approval survive restart. Running again after completion prints the existing receipt. L4: first run promotes v2 from 2/4 to 4/4 on a tiny held-out fixture set. Later runs keep v2 because the same candidate gives no further gain. L5: opens one local investigation linking export complaints with a release; remembers that the cause is unconfirmed. A model-proposed refund pauses for escalation and this service never executes it. No scheduler is supplied. L3 creates l3-records.sqlite and l3-checkpoints.sqlite. L4 creates l4-skills.sqlite. L5 creates l5-memory.sqlite, but execution checkpoints are in memory. Delete the local database files to reset the demos. L0-L2 use process-local state. Scope and limitations: - L0 is intentionally not an agent. L1 is assistance; L2 is a fixed team workflow. - L3-L5 show selected mechanisms, not complete implementations of those levels. - The L4 candidate is a choice between trusted code versions. No generated code is executed. Four evaluation fixtures are far too few for production claims. - A checkpoint does not supply identity, authorization or remote exactly-once writes. Real connectors need scoped service identities, target idempotency or a transactional outbox, outcome checks, concurrency controls and audit records. - L5 is a constrained research sketch with one permitted inquiry adapter, one internal action and a fixed hypothesis-memory schema. Its fixture question is scripted. Live question selection does not prove general autonomous discovery. - The L5 relationship check supports opening an investigation; it does not establish that a release caused a defect. Evidence content remains untrusted. - Production learning needs protected evaluations, broader outcome checks, versioned releases, rollback, and maintenance ownership. The canonical source is content/editorial/levels-of-ai-native-organizations/. Published downloads are generated byte-for-byte by scripts/autonomy_figures.py.