Core Concepts
CatalystRL is built on three keystones that work together to create agents that improve themselves over time.
Trust Engine
Progressive capability unlocks based on demonstrated reliability. Agents earn trust through successful operations and lose it through failures.
- Four trust levels: LOW, MEDIUM, HIGH, AUTONOMOUS
- Capability-gated operations
- Trust score persistence across sessions
- Automatic trust updates based on outcomes
Evolution AI
Controlled experiments on skill configurations with statistical winner selection. Zero-downtime deployment of improvements.
- A/B testing for skill variants
- Statistical significance detection
- Automatic rollout of winners
- Rollback on performance regression
ABES
Autonomous Bounty Evolution System. Skills detect their own failures and create structured improvement requests that can be resolved without human debugging.
- 9 bounty principles (P1-P9)
- Pattern-based error detection
- Evolver agents for resolution
- Human-in-the-loop for complex issues
How They Work Together
Skill executes an operation
Trust Engine checks if the skill has permission for this operation type.
Operation succeeds or fails
Trust score updates based on outcome. Evolution AI records metrics.
On failure: ABES creates bounty
Pattern detected, context gathered, bounty created for evolver or human.
Bounty resolved, skill improved
Fix applied, skill retested, trust score recovers with successful runs.
Additional Concepts
Gates
Safety checkpoints that validate operations before execution. Four types: Processing Integrity, Escalation, Constraint, and Permission Recovery.
Memory Architecture
Four-layer memory system: Global (Claude), Cross-Session (~/.claude/memory), Repo (CLAUDE.md), and Session (ephemeral).
Skill Families
Related skills grouped under orchestrators. Examples: commit family (scan, stage, create, push), shutdown family (memory, services).
Offline First
All skills work without network. Platform features queue operations for later sync. No cloud dependency for core functionality.