Skip to content

Latest commit

 

History

History
35 lines (28 loc) · 2.89 KB

File metadata and controls

35 lines (28 loc) · 2.89 KB

Roadmap

v0.1.0 (launch)

  • Vendored, tested diagram engine (architecture, workflow, sequence, dataflow, lifecycle)
  • Knowledge base: 12 grounded reference documents plus decision trees and source provenance
  • Skill definition: enterprise architect persona, clarification loop, framework decision tree, production and interview modes
  • Agent-native diagram vocabulary (llm-router, retriever, vector-store, guardrail, eval-loop, tool, memory-state, queue, human-review, model-gateway)
  • Self-contained design document template with embedded diagrams
  • Three gallery examples: enterprise support agent, legal RAG document assistant, autonomous coding system (interview mode)
  • Dual-theme diagram previews in the README (PNG via headless Chrome, plus a standalone-SVG exporter)

Next

  • Run the skill against a wide set of unscripted prompts and tune the clarification loop from what it asks (and fails to ask)
  • Knowledge base refresh automation: link checking plus a staleness report per source
  • Design diffing: compare two generated designs for the same use case (agentify diff)
  • Export to ADR (architecture decision record) format (agentify adr)
  • Model-graded eval to complement the current substring and rung checks
  • Additional diagram views: orchestration topology, eval pipeline, cost breakdown
  • Community example gallery with a submission workflow

Not planned

Recording declined ideas so contributors find the rationale rather than reopening them.

Idea Why not
A hosted service or web app Agentify generates a self-contained file you own. A backend adds an account, a data-retention surface, and an attack surface for zero gain over "run the skill, get an HTML file."
Running the systems it designs Agentify is a design tool, not an agent runtime or orchestration framework. It tells you what to build and why; building and running it is a separate job with mature tools already.
Auto-layout for diagrams Inherited from the engine's design: the semantic placement (a guardrail bracketing the model, the identity boundary around the agent) is the value. Auto-layout flattens that into a generic grid.
A diagram editor UI Positioning is generator plus viewer, not editor. Edit the JSON IR and re-render; the renderers are deterministic.
Replacing security, legal, or compliance review A generated design grounds decisions and flags risks; it does not sign off on them. A human owns the final call, and the design says so.
Baking in model names and prices These go stale in weeks. They are pulled live at generation time and dated, never committed to the knowledge base.
A giant "do everything" prompt instead of the knowledge base The bundled, cited knowledge base is the moat and the auditability. Folding it into one prompt loses the provenance and the ability to review and extend sources independently.