ERL Heuristics Pool
Experiential Reflective Learning. The swarm remembers what worked and references past lessons in future cycles.
Why
A single agent learns nothing across conversations. A single cycle teaches nothing to future cycles. Without persistent reflection, each improvement is amnesia.
ERL solves this: after every improvement cycle, the responsible agent writes a heuristic to output/swarm_architect/heuristics.md. Future cycles read this file before proposing new changes.
Structure
Each heuristic captures:
- ●Context: what problem triggered this lesson
- ●Action taken: the specific fix or improvement
- ●Outcome: what happened, verified or not
- ●Generalization: the rule-of-thumb derived from this experience
Example Heuristics
From the current pool:
When an agent ignores a rule for 7+ cycles despite repeated prompts, escalate from text rules to programmatic verification (shell-enforced). — Derived from github_scout B-013 (text rules failed 7 generations, 8th-gen used shell validation).
Content pipeline stability improves when LLM orchestrates once and a fat skill handles all sub-steps. Multi-step LLM orchestration of deterministic scripts is fragile. — Derived from content_publish pipeline.sh migration.
Integration
Every improvement cycle starts by reading heuristics.md. swarm_architect checks: "Have we seen this failure mode before? What did we learn?"
Related
- ●SAGE Critic Framework — heuristics feed future Critics
- ●Improvement Cycles — where heuristics are applied
- ●Self-Evolution Agents — who maintains the pool