Powerful agent frameworks exist, but they are often framework-heavy, cloud-centric, or opaque. Many teams want a local, auditable loop: a pattern they own, that improves with use, without shipping data to a vendor.
- Pattern compiler — Goals plus 2–5 example traces become a structured LoopSpec (prompts, tools, success checks, limits), versioned on disk.
- Execution kernel — ReAct-style plan → act → observe → score → adapt with hard caps (steps, time, remote usage).
- Trust boundary — Default local-only. Optional Claude or n8n connectors require explicit configuration and environment credentials.
PatternLoop is not “swap the model and rerun the same prompt.” The product is the loop artifact and the kernel that runs it safely.
- Replacing LangGraph/CrewAI for every enterprise workflow.
- Hosted multi-tenant SaaS inside this repository.
- Sending user content to the cloud without an explicit, named connector.
| PatternLoop | Typical agent framework | |
|---|---|---|
| Primary artifact | LoopSpec on disk | Code + graph definitions |
| Default runtime | Ollama locally | Often cloud APIs |
| Learning | Compile examples → LoopSpec; adapt within bounds | Often static prompts or separate training |
| Safety posture | Allowlists, redacted logs, opt-in remote | Varies |
Local agent OS: compile your workflow into a loop that improves until the goal predicate passes.