AgentFlow is a DSL for AI agents democratizing non-developer tool, workflow and UI creation in the spirit of tools like Excel Macros, Tableau, and Zapier.
AI agents running everything. Code running across companies handling various tasks: answering customer service requests, reading documents, doing analysis, completing applications, completing work and projects that weren't possible to get done before.
- Is Python the only way? Will it scale and be maintainable for the vision of scores of agents helping with work?
- How do non-developers contribute? analysts, project managers, power users, quality assurance people, SME's
- Contributions to prompts, instructions and evals? These are the new operating procedures and knowledge assets of a company.
- These assets will move out of code just like views/frontend separated from PHP style single file spaghetti when MVC/RoR conventions took off.
- Will they live in Prompt Repos or user friendly CMS tools for changement management?
- Prompts are not taught as structured components or managed this way, but can be broken down like: system, groundings, instructions, examples, outputs, guardrails
- AI needs patterns and solution architecture not just vibe code and gen code fast.
- LLM, AI and agents shouldn't just be API calls, Python scripts, code without conventions, and the latest framework
- There is a an AI ecosystem vocabulary that will emerge and reach a wide class of AI workers, it may change in name. Today's its: prompts, intents, goals, workflow, metrics, memory, context, evals and HTIL with knowledge extraction for model/context/RAG improvement alongside energy/compute efficiency.
- We need a lot more investment in end user experience and human computer interface.