The level is the single knob that ties together write-capability and
gate policy — it is not the same as a tool permission mode (that is
harness-permissions, which the engine derives from the level).
L1 — Report-only. Discovery and visibility, no automated fixes.
The maker runs read-only; every proposal is escalated to a human as
a report. The safest place to start any loop.
L2 — Assisted. The maker may write inside an isolated sandbox,
but a human gates every change before it lands. The default for a
loop you are actively building trust in.
L3 — Unattended. Runs within strict guardrails: only actions on
the gate’s allowlist commit automatically; anything else still
escalates. Reserve for narrow, well-understood loops.