Discovery Propagation
The system should learn from what it discovers.
The Pattern
Section titled “The Pattern”When an agent discovers a more efficient approach, a missing capability, or an insight useful beyond the current task, it proposes (or adopts) that improvement for future use.
The system evolves through use, not just through deliberate redesign.
What Gets Propagated
Section titled “What Gets Propagated”Tool Improvements
Section titled “Tool Improvements”- A skill that could be more efficient
- An MCP server missing a useful capability
- A prompt pattern that worked better than expected
Environment Insights
Section titled “Environment Insights”- A convention that should be standardized
- A shortcut others would benefit from
- A failure mode worth guarding against
Knowledge Capture
Section titled “Knowledge Capture”- Context that future tasks will need
- Decisions worth remembering
- Patterns worth encoding
Implementation
Section titled “Implementation”- Detection: Agent notices something worth propagating
- Proposal: Surfaces the finding (don’t silently mutate)
- Review: Human or process evaluates the proposal
- Adoption: Improvement is encoded for future use
- Propagation: Change reaches the appropriate scope (team, org, etc.)
Guardrails
Section titled “Guardrails”- Propose, don’t silently adopt (humans approve changes)
- Scope appropriately (not every insight is org-wide)
- Avoid churn (filter for meaningful improvements)
- Track provenance (where did this come from?)
The Vision
Section titled “The Vision”A self-improving system where:
- Good patterns spread automatically
- Friction gets reported, not endured
- Agents contribute to their own tooling
- The system gets better through use
Anti-patterns
Section titled “Anti-patterns”- Insights that die with the session
- Silent mutations without review
- Over-propagation (noise drowns signal)
- No mechanism to receive proposals