Field Reference · Newsletter Deep Dive · August 2026

The Developer's Rosetta Stone.

When you switch AI platforms, the model is the easy part. The friction is vocabulary — the same capability wearing a different name in every stack.

Every team that adopts a second AI platform rediscovers the same truth. The hard part of the migration is almost never the model, the latency, or even the pricing. It's the dictionary. The thing you call a tool in one stack is a function in another, a skill in a third, and a connector in a fourth. The "memory" you configured in one console is a "knowledge base" in the next and a "retrieval strategy" in the one after that. You already understand the concepts. You're just being forced to relearn their names — and that relearning is where weeks quietly disappear.

This guide is built for developers, systems architects, and technical builders who move between vendors and need to translate fast. It lays down the canonical nomenclature for each platform, then puts the players in each category side by side so you can map your existing mental model onto a new vendor's vocabulary and start building the same afternoon.

The premise

Switching platforms isn't learning new capabilities. It's learning new names for capabilities you already have. This is the dictionary that collapses that translation cost to near zero.

The landscape sorts into four categories. Enterprise agent runtimes host and govern autonomous agents at scale. Agentic coding environments write and run code inside your repositories. Data-native app builders bridge enterprise data lakes to interactive applications. And the skills & extension layer packages domain workflows so a generalist model becomes a specialist. Each section below opens with a quick-reference map, then goes in-depth on where the vendors genuinely diverge.

Lime = the universal concept — the translation key.
Neutral cells = each vendor's dialect for that same concept.
01

Enterprise Agent Platforms & Cloud Runtimes

Players: Google Gemini Enterprise · AWS Bedrock + Bedrock AgentCore · Microsoft Foundry / Foundry Agent Service

These platforms exist to host, execute, secure, and scale autonomous agents grounded in enterprise data. They look wildly different in their consoles and share almost no terminology — but structurally they converge on the same five building blocks.

The concept translation map
Google Gemini Enterprise
Builder canvasAgent Studio
Code frameworkAgent Development Kit (ADK)
MemoryMemory Bank
Grounding / RAGRAG Engine
Gateway / governanceSemantic Governance
AWS Bedrock AgentCore
Builder canvasAgentCore Harness
Code frameworkStrands / Code-based Loop
MemoryAgentCore Memory
Grounding / RAGAgentCore Memory
Gateway / governanceAgentCore Gateway
Microsoft Foundry
Builder canvasFoundry Portal
Code frameworkFoundry Agent Service
MemoryFoundry IQ
Grounding / RAGFoundry IQ
Gateway / governanceControl Plane / Content Safety
Five universal concepts · three dialects · read each row across the columns
Nomenclature cheat sheet
Low-code / no-code canvas & config-driven builder
Gemini Enterprise
Agent Studio
Bedrock AgentCore
AgentCore Harness (Preview)
Microsoft Foundry
Foundry Portal / Templates
Code-based agent framework & custom loop orchestrator
Gemini Enterprise
Agent Development Kit (ADK)
Bedrock AgentCore
Strands / Framework Agents
Microsoft Foundry
Foundry Agent Service
Scalable, managed container / serverless host runtime
Gemini Enterprise
Agent Engine
Bedrock AgentCore
AgentCore Runtime
Microsoft Foundry
Agent Service Endpoint
Conversational vector / document memory bank
Gemini Enterprise
Memory Bank
Bedrock AgentCore
AgentCore Memory (Long-term)
Microsoft Foundry
Foundry IQ
Secure API integration / tool registry
Gemini Enterprise
Skill Registry
Bedrock AgentCore
AgentCore Gateway / Tools
Microsoft Foundry
Foundry Tools / AI Services
Trust & safety guardrails / gateway routing
Gemini Enterprise
Agent Gateway / Semantic Governance
Bedrock AgentCore
AgentCore Gateway
Microsoft Foundry
Control Plane / Content Safety
Automated LLM output verification & monitoring
Gemini Enterprise
Agent Evaluation (Simulated / Offline)
Bedrock AgentCore
AgentCore Evaluators
Microsoft Foundry
Foundry Evaluation & Traces
In-depth comparison

1 · Developer loop & local prototyping

Gemini Enterprise

Leans on Agent Studio — a visual web console — for fast prototyping, then transitions into code with the model-agnostic ADK. Console-first, code-second.

Bedrock AgentCore

Ships a state-of-the-art CLI (@aws/agentcore). Running agentcore dev spins up a local hot-reloaded environment with a web-based agent inspector and trace visualizer. CLI-first.

Microsoft Foundry

Code-driven from the start via AI Templates and standard inference SDKs (Python, C#, JS, Java), with VS Code extensions and Foundry Local to run LLMs on-device for free.

2 · Grounding & RAG

Gemini Enterprise

A dedicated RAG Engine links files to LLMs, built on auto-tuned Vector Search indexes. Retrieval is its own managed service.

Bedrock AgentCore

Grounding folds into AgentCore Memory, with retrieval strategies configured declaratively in agentcore.json. Memory and grounding are one surface.

Microsoft Foundry

Foundry IQ connects existing knowledge bases directly to agents — no manual chunking or vector setup. The abstraction hides the pipeline.

3 · Security, safety & governance

Gemini Enterprise

Agent Gateway inspects payload traffic; Semantic Governance configures granular behavior policies. Access is gated through Google Cloud IAM.

Bedrock AgentCore

AgentCore Gateway governs tool connectivity; AgentCore Identity bridges third-party keys and OAuth so agents act with scoped credentials.

Microsoft Foundry

Control Plane packs enterprise-grade Content Safety — red teaming, input/output filtering — plus deep audit tracing as a first-class feature.

02

Agentic Coding Environments, IDEs & Tools

Players: Google Antigravity · Claude Code · Cursor · OpenAI Codex

These are developer tools designed to write, test, debug, and run code directly inside repositories. They diverge most on two axes: how you inject persistent context, and how much autonomy you grant the sandbox.

Nomenclature cheat sheet
Multi-agent coordination & tool orchestration engine
Antigravity
AGY SDK
Claude Code
Agent SDK
Cursor
Composer 2.5
Persistent repository instructions & workspace memory
Antigravity
walkthrough.md
Claude Code
CLAUDE.md / Auto Memory
Cursor
.cursorrules / Semantic Index
Repeatable workflows or user-created slash commands
Antigravity
Task / Checklist
Claude Code
Skills
Cursor
Skills (/loop etc.)
Pre- or post-edit shell scripts triggered on actions
Antigravity
N/A
Claude Code
Hooks
Cursor
Cursor Automations
Isolated runtime to build, execute, and run tests
Antigravity
Terminal Sandbox
Claude Code
Routines / Sandboxes
Cursor
Shadow Workspaces / Cloud Agents
In-depth comparison

1 · Instructions & context injection

Claude Code

Reads CLAUDE.md from the project root at the start of every session to align with your styling guidelines. It also tracks Auto Memory — saving build commands, diagnostic paths, and tricks learned across sessions.

Cursor

Relies on .cursorrules files and builds a secure codebase index for full semantic search across the repo (invoked with the @Codebase symbol).

Antigravity

Uses Implementation Plans (implementation_plan.md), Tasks (task.md), and Walkthroughs (walkthrough.md) stored in the app data directory to enforce explicit phases: Research → Plan → Review → Execute → Verify.

2 · Execution sandboxes & autonomy

Claude Code

Supports Routines — running on managed Anthropic cloud sandboxes as asynchronous webhook/cron loops — alongside local terminal tool execution.

Cursor

Composer 2.5 runs tests and tracks compilations inside Shadow Workspaces. Agent Mode exposes a visual autonomy slider — autocomplete, inline edits (Cmd+K), or a fully autonomous Cloud Agent.

Antigravity

Operates a structured sandbox with gated shell-command authorization. Mutating state requires an explicit, user-approved handshake transaction before execution.

A note on OpenAI Codex. Codex sits in this category as a cloud-and-CLI coding agent, but it maps cleanly onto the same five primitives above — an orchestration engine, an instructions file, repeatable commands, action hooks, and a sandboxed runtime. Once you can read the columns, a new entrant rarely needs its own row; it just fills the existing ones with new names.

03

Data-Native & App Builders

Players: Databricks Agent Framework / Databricks Apps · Replit × Databricks

These tools bridge the gap between heavy enterprise data lakes and interactive, user-facing applications. The shared pattern is the same in both: an agent generates a UI that reaches a governed SQL warehouse through scoped, audited credentials.

Replit Agent
App template · writes the dashboard UI
Service Principal credentials
SQL Warehouse
Databricks compute engine
reads data
Databricks Workspace
Unity Catalog · governed lake

The data-to-app pipeline — an agent at one end, a governed catalog at the other

Nomenclature cheat sheet
No-code natural-language SQL querying agent
Databricks Agent Framework
Genie Spaces
Replit × Databricks
Replit Agent
Host container / runtime for interactive Python & JS UIs
Databricks Agent Framework
Databricks Apps
Replit × Databricks
Databricks App Template
Secure, structured database tool-call APIs
Databricks Agent Framework
Unity Catalog Functions
Replit × Databricks
Service Principal SQL Warehouse Connection
Audited, governed access router to data catalogs
Databricks Agent Framework
Unity AI Gateway
Replit × Databricks
Org Connectors (RBAC-gated)
Comparative architecture · two ways to ship
Path A · Databricks-native

Build inside the Agent Framework

  1. Register the tools. Expose secure SQL query tools as Unity Catalog Functions.
  2. Orchestrate. Design with a Supervisor Agent in Python, or start from a Knowledge Assistant template.
  3. Trace. Run and instrument execution through MLflow Tracing to log payload schemas.
  4. Deploy. Ship the UI directly onto Databricks Apps — serverless hosting, no separate infra.
Path B · Replit partner

Use the Replit × Databricks connector (Enterprise plan)

  1. Mint credentials. A Databricks admin creates a Service Principal — machine-to-machine client credentials.
  2. Register the connector. The Replit admin adds those credentials under Org Connectors and maps the SQL Warehouse hostname and HTTP path.
  3. Scaffold. A developer spins up an app from the Databricks App template.
  4. Let the agent build. Replit Agent writes the dashboard UI, reads metadata catalogs, generates secure queries, and deploys to a Replit staging URL — all from a natural-language prompt.
04

Skills & Extension Architecture

Players: Agent Skills / Open Skill Format · SkillsMP · Claude Code Plugins

Skills let developers package domain-specific workflows, API specs, utility scripts, and reference instructions — turning a generalist model into a specialized tool. The format is deceptively simple: a directory with a known structure that the runtime loads in layers.

The installed-skill directory
[Installed Skill Directory] 
├── yaml frontmatterLEVEL 1 · metadata · preloaded
├── SKILL.mdLEVEL 2 · instructions · on-demand
├── REFERENCE.mdLEVEL 3 · reference docs
└── scripts/ 
    └── execute_workflow.pyLEVEL 3 · executable · sandboxed
Comparative capabilities
Capability Agent Skills (API) Claude Code Skills claude.ai Custom Skills
How to install Set skill_id in the API container config Store directory in .claude/skills/ Upload a .zip package in UI settings
Discovery Auto-triggered via YAML description Auto-triggered via YAML description Auto-triggered via YAML description
Sharing scope Workspace-wide — shared API resource Personal or project — Git-trackable Individual only — no admin share
Network access Strictly blocked — local sandbox only Full access — inherits the user's machine Variable — gated by admin settings
Runtime dependencies Pre-installed packages only Can run local pip / npm packages Pre-installed packages only
The "three-level loading" performance trick

The reason a model can have hundreds of skills installed without drowning is progressive disclosure. Nothing heavy loads until it's actually needed — and the heaviest material never enters the context window at all.

01
Metadata≈ 100 tokens

Only the YAML name and description load at startup. This is what lets you install hundreds of skills with no meaningful token penalty — the model knows what each skill is for without reading how it works.

02
Instructionsloaded on match

The body of SKILL.md is read via bash and pulled into context only if the user's prompt matches the skill's description. Procedural detail arrives exactly when it becomes relevant — and not before.

03
Resources & scriptsstdout only

Executable scripts and heavy schemas run in the terminal; the model reads only the stdout. The raw code and reference guides never enter the context window — which is what keeps the loop fast and cost-effective even for large, complex skills.

The takeaway

Learn the columns, not the vocabularies.

The platforms in this guide will keep renaming things. New entrants will arrive with their own marketing terms for the same primitives. None of that should cost you a week ever again — because once you can read the rows, every "new" platform is just a fresh set of names laid over a structure you already understand: a builder, a framework, a runtime, a memory, a tool registry, a gateway, and an eval harness.

That's the real value of a Rosetta Stone. Not the translation of any single phrase — but the moment you stop translating and start reading.

The fastest engineers on a new platform aren't the ones who memorized its docs. They're the ones who recognized which familiar concept each new word was hiding.

Designing across these platforms? That's our work.

auxfirst designs agentic systems that stay coherent no matter which runtime, IDE, or skills layer they're built on — because good agent design lives above any single vendor's vocabulary.

Start a conversation →

Read next & further reading

The platform layer
Managed Agent Platforms — the build-vs-buy decision underneath these vocabularies Agent-First Design — designing the product these platforms run Agent-First Design Patterns — the reusable moves across stacks
The discipline underneath
What Is Agentic User Experience (AUX)? — delegation, trust and control The 10 AUX Heuristics — the evaluation cadence TrustKit & Trust Harness — testing agent behaviour across platforms
Where commerce meets the stack
Agentic Commerce Explained — MCP, A2A and the protocols agents transact through

Method · Capability names are taken from each vendor's own documentation at time of writing. Platform vocabularies change frequently; treat this as a translation aid, not a specification, and check the primary docs before building.


Emil Krzemiński is the founder of auxfirst, the agency for the agentic era. This guide is part of the auxfirst field-reference series — practical maps of the agentic toolchain for the teams building on it. For more, browse the newsletter archive or subscribe to the auxfirst Substack.