Agent Patterns: Learned from Run Logs
Purpose: Patterns extracted from agent runs using “thinking to summary” approach. Patterns help agents improve over time.
Related: RUN_LOG.md, AGENT_FEEDBACK_LOOP.md, CONTEXT_GRAPH_APPROACH.md
Pattern Confidence Levels
| Confidence | Criteria | Action |
|---|---|---|
| LOW | 1-2 examples | New pattern, track here |
| MEDIUM | 3-4 examples | Pattern is reliable, update agent PRD |
| HIGH | 5+ examples | Pattern is proven, auto-apply in agent |
Ticket Creation Agent Patterns
🆕 Pattern: Ticket Title Format (LOW Confidence)
Pattern: Remove “Linear Ticket:” prefix from ticket titles. Titles should be clear and actionable without the prefix.
Evidence:
- Run 1 (2026-02-05): Title had “Linear Ticket:” prefix → User feedback: “remove ‘Linear Ticket’”
Impact: Cleaner titles, better readability in Linear UI.
Action: Update agent to never include “Linear Ticket:” prefix in titles.
Confidence: LOW (1 example)
🆕 Pattern: Success Criteria Required (LOW Confidence)
Pattern: All tickets must include a “Success Criteria” section with checkboxes for clear acceptance criteria.
Evidence:
- Run 1 (2026-02-05): Missing success criteria → User feedback: “no success criteria”
Impact: Tickets are more actionable and testable.
Action: Update agent to always include success criteria section with checkboxes.
Confidence: LOW (1 example)
🆕 Pattern: Point Assignment Required (LOW Confidence)
Pattern: All tickets must include point assignment. Default: 1pt for small requests. 1pt = 1hr for increasing complexity.
Evidence:
- Run 1 (2026-02-05): Missing point assignment → User feedback: “no point assignment. default is 1pt for a small request, and then 1pt = 1hr for increasing complexity”
Impact: Better estimation and capacity planning.
Action: Update agent to always include points field with default 1pt for small requests.
Confidence: LOW (1 example)
✅ Pattern: Data Source Reference (LOW Confidence)
Pattern: Agent correctly references data sources (e.g., BigQuery tables) when creating data request tickets.
Evidence:
- Run 1 (2026-02-05): User feedback: “it referenced the data sources correctly, i wonder how this happened”
Impact: Tickets have better context, easier to execute.
Action: Reinforce this behavior in agent (keep doing this).
Confidence: LOW (1 example, but positive reinforcement)
Design-Ready Copy Agent Patterns
✅ Pattern: Single Service → Service 2-pager (MEDIUM Confidence)
Pattern: Single-service campaigns consistently use Service 2-pager archetype.
Evidence:
- Run 1 (2026-02-04): insurance-broker-lead-intake → Service 2-pager → Used successfully
Impact: Faster archetype selection, consistent output structure.
Action: Auto-suggest Service 2-pager for single-service campaigns.
Confidence: MEDIUM (1 example, but aligns with taxonomy rules)
Slack Deployment Worker Patterns
🆕 Pattern: Audience-Specific Message Versions (LOW Confidence)
Pattern: Slack deployment messages need different versions for different audiences. GTM audience needs non-technical, benefit-focused messaging. Engineering audience needs traditional PR review format.
Evidence:
- Run 1 (2026-02-05): Message created for GTM audience → User feedback: “it gave the appropriate context to a non-technical GTM audience, but there need to multiple versions of Slack messages depending on the audience. for example, my engineers would probably want a more traditional PR review and that looks different”
Impact: Messages are more effective when tailored to audience (GTM vs Engineering).
Action: Update worker to prompt for audience type and generate appropriate format (GTM deployment message vs PR review format).
Confidence: LOW (1 example)
✅ Pattern: Message Completeness (LOW Confidence)
Pattern: Slack deployment messages consistently include all required sections (demo link, PRD highlights, process replaced/created, where it lives, context graph evolution, feedback CTA).
Evidence:
- Run 1 (2026-02-05): All completeness checkboxes met → Used as-is
Impact: Messages are complete and ready to send without edits.
Action: Reinforce this behavior (keep doing this).
Confidence: LOW (1 example, but positive)
Pattern Notes
- Patterns are extracted using “thinking to summary” approach from learn-extraction skill
- Patterns move from LOW → MEDIUM → HIGH as evidence accumulates
- When pattern reaches MEDIUM confidence, create PR to update agent PRD/taxonomy
- When pattern reaches HIGH confidence, auto-apply in agent logic