Continuous learning
Improvement comes from memories, retrieval, skills, workflows, agent selection, orchestration, prompts, policies, tools and evaluations — not from model weights.
Learning loop
Task ↓ Execution ↓ Outcome ↓ Feedback / Evaluation ↓ Reflection ↓ Candidate Learning ↓ Evaluation ↓ Promotion ↓ Better Future Execution
Offline learning / “sleep”
Execution Traces
↓
Offline Consolidation
↓
Pattern Discovery
↓
Candidate Memory / Skill / Workflow
↓
Evaluation
↓
Promotion or RejectionMerge duplicates, drop noise, detect recurring failures and successful patterns, generalize experience, propose skills, discover workflows and automation opportunities.
Evaluation & promotion
Candidate Improvement
↓
Evaluation Suite
↓
Compare with Production Version
↓
Better + Safe?
├── Yes → Promote
└── No → Reject / ReviseLearning must not automatically become truth. Deterministic rules, tests, historical replay, golden datasets, human review, LLM judges, adversarial and security tests, cost and latency.
Learn → automate
Record how work is actually done.
Help humans perform parts of the process.
Propose actions and workflows.
Execute behind approval gates.
Automate low-risk workflows within policy.
Validated workflows evolve from experience.
First vertical slice
Repeated executions produce traces. Traces reveal a recurring pattern, which becomes a candidate workflow, replayed against history, evaluated, approved by finance and promoted to v1. Later traces expose variants and exceptions — producing validated v2 and v3.
Invoice received → extract fields → supplier lookup → PO match → VAT / policy checks → anomaly review → approval → ERP draft → archive