A year ago (May 2025), I saw that each AI coding tool (Cursor, Cline, WindSurf) had their own proprietary way of locally configuring system prompts and rules. They were similar, but not standardized. Similarly, consumer AI tools (ChatGPT, Claude, etc) were trickling-out features like "Projects" and their own versions of agent configurations - each operating differently.
As we AI early adopters test out these features (and even try to develop some of them ourselves), we started to see ourselves porting use cases implemented in one tool over to other tools with similar capabilities. But, this use-case portability problem just seems to be proliferating as tools race to add features, and industry standards lag behind innovation velocity.
Now, as I am embarking on some larger time investments into automating large-domain use cases, I don't want to get trapped with single-tool / single-framework lock in. I want to be able to build using industry standards (as they evolve) and maximize my ability to port use case solutions across tools. For example, if I build a market research tool or a work decomposition module (or better yet, an end-to-end software development automation pipeline), I want to be able to port it to any current or future AI platform. I want to try it in Kiro, test it in Anti-Gravity, try it with Claude Code and port it to (any)-Claw.
To do this, I decided to do some research on the current state of evolving standards in the AI space with regard to agent definitions, skills, tools, workflows, guardrails, and the like. As with most research I perform, I first iterate over an ontological framework to organize my thinking (and that of my AI tools), so that I can get a wide aperture and narrow in on key domains of research. From this work, here is the set of tier-1 domains that I found useful in evaluating standards for AI-tool maturity, standards, interoperability and solution fungibility.
Ontology-Based Approach to AI Standards
When assessing the field of AI Standards, we can use an ontology-based approach to categorize the topics of concern. Here are 12 distinct concern areas that tools must address to configure, encode, and orchestrate intelligent agents.
01 - Identity
Agent identity & persona: Role, name, goals, backstory, tone. How an agent presents itself and constrains its character across sessions.
02 · Skills
Skill / capability definition: Modular, reusable packages that encode specialist knowledge, tool procedures, and output formats — loaded on demand.
03 · Tools
Tool integration protocol: Standardized channel for agents to invoke external tools, APIs, file systems, and live data sources.
04 · Orchestration
Inter-agent communication: How agents discover peers, delegate tasks, and exchange structured results across framework or vendor boundaries.
05 · Hooks
Lifecycle event hooks: Callbacks at spawn, pre/post tool use, prompt submit, and stop. Enables audit logging, approval gates, and auto-formatting.
06 · Memory
Memory architecture: Working (context window), episodic (session), and long-term (persistent) memory tiers with defined retrieval semantics.
07 · Identity
Agent manifest / config: Machine-readable declaration of agent identity, allowed tools, resource access, permitted paths, and model selection.
08 · Routing
Model selection & routing: Per-agent or per-task assignment of LLM provider/model. Enables cost/quality tradeoffs across a multi-agent workflow.
09 · Ops
Observability & telemetry: Structured traces, cost attribution, token accounting, and audit logs across agent invocations and tool calls.
10 · Safety
Guardrails & safety constraints: Output filtering, permission scoping, content policies, approval gates, and harm prevention at the infrastructure layer.
11 · Workflow
Spec-driven development: Requirements, design, and task documents as durable first-class context — planning phase separated from execution.
12 · Knowledge
Knowledge & RAG integration: Retrieval-augmented context from codebases, docs, or vector stores — indexed, scoped, and injected at query time.
