Rulesmith — Development skill for Claude Code

Agent Framework

A lightweight, reusable framework for building LLM agents: tool use, memory, and human-in-the-loop, with no orchestration library dependencies.

Problem Statement

IEP (Intelligent Event Processing) systems route events to automated actions via rules. Rules age badly: new event types may emerge, operational patterns shift, and some events may fall through, operators silently by handling unmatched events by hand: raising alerts, restarting services. Those interventions never turning into rules.

Finding these gaps requires three things happening together: knowing which event types consistently go unmatched, understanding how the ops team has historically responded to those events, and translating that pattern into a correctly-scoped rule. If done manually, this is a multi-step investigation across different data sources. Done wrong, it produces rules that are too broad (false positives) or too narrow (miss real incidents).

**RuleSmith solves this by acting as an analyst agent.** It is given a goal — *find events with no matching rule in the last 24 hours and suggest rules for them* — and it autonomously gathers the evidence, reasons over it, proposes a rule with a data-backed rationale, and then waits for human approval before writing anything to the system. The operator reviews one clearly-reasoned proposal rather than doing the investigation themselves.

---

Demo

[Watch the 5-minute demo](https://sap-my.sharepoint.com/:v:/p/ahad_abbas_rizvi/IQCYaZ7F7j2ES6L49rC8baS6AdALNKT5Q85jFHibmo0YPjg?e=9SMrVC)

---

What it does

The framework gives you a **free ReAct loop**: the LLM decides which tools to call, how many times, and in what order. You define tools with a decorator, hand them to an `Agent`, and call `agent.run(goal)`.

---

5 Capabilities

	Capability	Where
1	Tool Registry — @registry.tool decorator turns any Python function into a callable tool with an auto-generated OpenAI schema