My view of AI
I approach AI as a tool for concrete work. It can draft text, analyse documents or prepare a workflow. The decisive question is what someone else can then do with the result. Does it fit the task? Are sources and limits visible? Who makes the next decision?
An agent adds tools and a workflow to a language model. Requirements change: an error in wording differs from a database change. Task descriptions, access, review and recovery therefore belong in the design.
My projects are starting points for these questions. Alpine-X explores bounded workspaces and approvals; other experiments are in the directory. These articles distinguish possible architectures from demonstrated results. One successful example does not establish a reliable success rate.
Three questions to go further
- When is an agent useful?
- When steps, tools and feedback interact and a fixed workflow alone is insufficient. For a predictable transformation, a simple script may be easier to understand and operate. Start the comparison with the task.
- How do I recognise a good result?
- Through predefined criteria and realistic tests, including missing information, conflicting documents and technical failures. Readability and correctly completed work must be evaluated separately.
- What remains the human role?
- People set goals and limits, judge disputed cases and take responsibility. Approval helps only if the reviewer can see essential information and still effectively prevent an action.
Articles & practical guides
Evaluating AI agents: from a good impression to a useful result
An agent can report convincingly while doing the wrong work. A framework for checking tasks, results, permissions and repeatable trials.
AI agents beyond the chat window
A chat window can help with ideas, but workflows need more. Why operational AI needs a workspace and a person to approve its actions.
Projects as a starting point
Sacretus
Sacretus: connecting documents, validated facts and source versions. A Stoicera research project involving Raphael Lugmayr. Seeking pilot partners.
Stoicera Bid Desk
Stoicera Bid Desk screens tenders against a company profile: requirements, supporting documents and deadlines with sources. A project involving Raphael Lugmayr.
Alpine-X AI
Alpine-X: AI agents with their own workspaces, defined workflows, documented steps and human review. By Raphael Lugmayr.