Lab repo for Engineering Reliable Agentic AI Systems (Packt workshop, 29 August 2026). TicketCloser on a sample CRM.
This repository also configures its own agents. See what reliable-agentic-lab tells them →
RichardHightower/reliable-agentic-lab
Agent Claude Code
Investigates a ticket and the target app, and drafts an updated ticket body that fills its missing fields. Holds no write tool of any kind; its draft is the text of its final reply, for the enhancer-loop orchestrator to write to a candidate file and judge before it touches any real file.
RichardHightower/reliable-agentic-lab
Agent Claude Code
Reads one ticket file and reports which required fields for its kind have real content. Never writes anything, never grades its own draft.
RichardHightower/reliable-agentic-lab
Agent Claude Code
Writes the code that makes a failing test pass. Writes under app/ and nothing else, and cannot reach the tests.
RichardHightower/reliable-agentic-lab
Agent Claude Code
Scores one attempt against the rubric and reports whether the ticket is done. Holds no write tool.
RichardHightower/reliable-agentic-lab
Agent Claude Code
Turns one ready ticket into an ordered list of steps, one test step and one code step per acceptance criterion. Writes steps.jsonl and nothing else.
RichardHightower/reliable-agentic-lab
Agent Claude Code
Writes the failing test for one acceptance criterion. Writes under tests/ and nothing else.
RichardHightower/reliable-agentic-lab
Agent Claude Code
Returns Mermaid or PlantUML source for one figure, and simplifies it when the harness reports what the render lost. Holds no shell and no write tool.
RichardHightower/reliable-agentic-lab
Agent Claude Code
Scores a finished white paper against a fixed rubric and reports whether it is done. Holds no write tool.
RichardHightower/reliable-agentic-lab
Agent Claude Code
Answers one research question from primary sources and returns the atomic claims it found, each attached to a source and a quote. Writes nothing.
RichardHightower/reliable-agentic-lab
Agent Claude Code
Independently checks one claim against a source it finds itself. Never sees the original research. Writes nothing.
RichardHightower/reliable-agentic-lab
Agent Claude Code
Writes one section of the white paper from verified claims only. Writes under sections/ and nothing else. Assembly owns paper.md.
RichardHightower/reliable-agentic-lab
Agent Claude Code
Repairs the application code until a failing suite passes. Writes under app/ and nothing else, and cannot reach the tests.
RichardHightower/reliable-agentic-lab
Agent Claude Code
Scores one repair attempt and reports whether the branch is genuinely fixed. Holds no write tool.
RichardHightower/reliable-agentic-lab
Agent Claude Code
Returns a chart spec from planned figures and data tables. Holds no write tool.
RichardHightower/reliable-agentic-lab
Agent Claude Code
Extracts the facts, numbers, decisions, and terms from one finished section. Holds no write tool.
RichardHightower/reliable-agentic-lab
Agent Claude Code
Scores a white paper outline against flow, completeness, titles, and corpusfit. Holds no write tool.
RichardHightower/reliable-agentic-lab
Agent Claude Code
Turns one topic into a two-level white paper outline. Runs at the start of a run. Writes nothing.
RichardHightower/reliable-agentic-lab
Agent Claude Code
Grades one section against its outline row. Holds no write tool.
RichardHightower/reliable-agentic-lab
Agent Claude Code
Edits a judged white paper outline against the judge's objections, changing only what was named. Holds no write tool.
RichardHightower/reliable-agentic-lab
Agent Claude Code
Proposes the search domains for one topic. Python admits them. Holds no write tool and no search tool.
RichardHightower/reliable-agentic-lab
Agent Claude Code
Finds the public page that carries one cabinet source, by open web search. A cross-reference, not research. Writes nothing.
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: