Autonomy

Self-improvement

Brigade can get better over time — but never behind your back. Behavior changes flow through a human-gated, reversible loop. Nothing that affects how your crew behaves ships without an explicit approval, and every applied change can be rolled back.

Two lanes#

There are two distinct improvement lanes, and only one is autonomous:

• Lane A — facts. The nightly consolidate/reflect pass lets memory facts evolve on their own (confirm repeated beliefs, merge duplicates, evict noise). See Memory (Tideline).
• Lane B — behavior. Changes to preferences, skills, or prompts are deliberately not autonomous. They go through the gated loop below.

The behavior loop#

1. 1

   Propose

   Telemetry — the memory feedback and event log — drives a proposal: a concrete diff against a preference, a skill, or a prompt, with a rationale.
2. 2

   Evaluate (optional)

   A held-out evaluation can gate the proposal before it ever reaches you, so weak changes are filtered out automatically.
3. 3

   Approve

   You review and approve. Until then the proposal sits pending and changes nothing.
4. 4

   Apply, reversibly

   On approval the change applies — and stays reversible. If it regresses, revert it.

Safety constraints#

A configurable read-only set can never be the target of a proposal — proposals against it are dropped and apply refuses. That keeps core identity and the things you have marked off-limits out of the self-improvement path entirely.

Skill & behavior review#

Alongside the proposal loop, Brigade can review its own behavior and curate skills from usage — surfacing stale or unused skills and reviewing how it is doing its work. These reviewers run on a schedule and can be tuned or disabled through environment settings; like everything in Lane B, what they surface is a suggestion for you, not an automatic change.