Operator Patterns¶
This page describes practical ways to run Kabot well over time.
Pattern 1: One Stable Test Prompt Set¶
Keep a small set of known-good prompts for: - temporal response - filesystem navigation - model override - dashboard chat - one inbound channel message
Run them after meaningful config or runtime changes.
Pattern 2: Separate Observation From Mutation¶
In the dashboard and gateway world, first observe, then mutate.
Good sequence: 1. open dashboard 2. inspect health, sessions, and nodes 3. confirm scope/auth mode 4. then use write actions
Pattern 3: One Change At A Time¶
Do not change: - model chain - channel binding - gateway auth token - memory mode
all at once.
Change one thing, smoke test, then continue.
Pattern 4: Scope-Aware Tokens¶
Use:
- operator.read for observation-only operators
- operator.write only where mutation is necessary
- ingress.write only where webhook ingestion is needed
Pattern 5: Role-Based Agents¶
Use multiple agents when roles genuinely differ.
Examples: - support agent for customer-facing channels - research agent for long-form exploratory tasks - work agent for coding and operations
Pattern 6: Treat Skills As Capabilities, Not Magic¶
A skill still needs: - relevant env keys - any required binaries - a good runtime path - realistic operator expectations
Pattern 7: Use Doctor And Smoke Checks Proactively¶
Operator confidence comes from repeatable checks, not intuition.