Mail-driven research notes
GroupForge Wiki
Curated source notes from the Copilot Cowork pipeline, shaped into a calm browsable wiki with search, clear status labels, and room for the collection to grow.
46public pages
4recommended items
14protected ideas
2026-05-20 18:10latest publish
Public Pages¶
moderate confidence
Agent-to-Agent (A2A) Protocol
The Agent-to-Agent (A2A) protocol is an open standard for communication and interoperability between AI agents built by different vendors, running on different frameworks, and deployed across different cloud environments. Where the Model Context Protocol (MCP) standardises how an agent connects to its *tools*, A2A standardises how agents connect to each other — enabling one agent to delegate tasks to, query, and hand off work to another agent it has never directly integrated with before.
high
Agent Skills authoring — overview
This is a vendor-neutral, deeply-worked authoring guide for **Agent Skills**: file-based, on-demand capabilities that the major AI coding agents — Anthropic Claude, GitHub Copilot, and OpenAI Codex — all now read from the same `SKILL.md` format. The guide is written so that someone with no prior Skill experience can plan, build, test, and distribute a portable Skill that runs across all three platforms with only narrow vendor-specific tweaks.
high
Agent Skills — distribution and sharing
A Skill that lives only on the author's laptop has done a fraction of its job. This page covers the distribution end of the lifecycle: how to package a Skill so others can install it, what the per-vendor install flow looks like in practice, how the plugin and marketplace wrappers work, and how organisations roll Skills out at scale.
high
Agent Skills — planning and design
This page covers the *before-you-open-the-editor* phase of authoring a Skill: choosing a candidate use case, defining success criteria for it, gathering technical requirements, drafting the directory and `SKILL.md` skeleton, and — most important — writing a `description` field that triggers reliably without spilling into other Skills' territory. Everything here is vendor-neutral; the [vendor-extensions sub-page](agent-skills-vendor-extensions) covers fields specific to Claude, Copilot, and Codex.
high
Agent Skills — testing and iteration
A working Skill has to clear three hurdles in order: the agent must **trigger** it on the right turns, the agent must **execute** it correctly when triggered, and the cost / latency of triggering and executing it must stay within a budget the user is willing to pay. This page covers how to test all three, the signals each one gives off when it fails, and the iteration moves that fix the most common failure modes.
high
Agent Skills — vendor extensions reference
The Agent Skills open specification defines a **small, portable core** — two required frontmatter fields (`name`, `description`) and a handful of optional ones (`license`, `compatibility`, `metadata`). The Anthropic, GitHub, and OpenAI implementations all read that core identically, which is what makes `SKILL.md` cross-vendor in the first place. Each vendor then adds its own optional extensions for things the core does not cover — tool allow-listing, picker UI hints, invocation policy, MCP dependencies, execution context isolation. The extensions do not interoperate.
moderate confidence
Agentic RAG
Agentic RAG extends standard Retrieval-Augmented Generation by replacing the fixed one-shot retrieval pipeline with a dynamic loop driven by an AI agent. Where standard RAG retrieves once and generates once — a librarian who fetches one book and hands it over — Agentic RAG uses an agent as a researcher who decides *what* to retrieve, *from where*, *whether the result is good enough*, and *whether to try again* before generating a final answer.
high
Coding agents — side-by-side comparison
By May 2026 the agentic-coding category has settled into a stable quartet of platform-scale offerings: Anthropic's **Claude Code** (with the underlying Claude Agent SDK), Microsoft's **GitHub Copilot** family of agent surfaces, OpenAI's **Codex** stack, and Google's **Antigravity** platform. Each one ships a multi-surface, multi-trigger experience that can read a repository, edit files, run shell commands, drive a browser, and open a pull request without continuous human steering. Each one wraps a frontier reasoning model inside a sandbox, a permissions model, a configuration file convention, and an MCP-based tool extension surface. They look superficially similar and they all read parts of each other's conventions — yet at the level of execution model, security posture, model lineup, pricing, and target workflow the differences are substantial enough to drive real procurement decisions.
high
Coding agents — the 2026 landscape
By the middle of 2026 "agents" has stopped being a buzzword and become a product category. Four vendors — Anthropic, GitHub/Microsoft, OpenAI, and Google — each ship a multi-surface stack of autonomous coding tools that share a recognisable shape: a model in a tool-use loop, a project-level instructions file, a Model Context Protocol bus to external tools, a sandbox of some kind, and a way to delegate work either to your editor or to a remote VM. The vendor stacks differ in detail and in philosophy, but the underlying primitives have converged to a degree that was not obvious twelve months earlier. This page is the hub that ties the four vendor-specific pages — [claude-agents](claude-agents), [github-copilot-agents](github-copilot-agents), [codex-agents](codex-agents), and [antigravity-agents](antigravity-agents) — into a single mental model of the category. A more granular side-by-side lives on [agents-comparison](agents-comparison); this page is the umbrella narrative.
moderate confidence
Securing Autonomous AI Agents on Kubernetes
## Why AI Agents Break Traditional Kubernetes Security Models
wiki note
Superseded
Curated wiki note.
high
Google Antigravity
Google Antigravity is the agentic development platform Google launched on **November 18, 2025**, alongside the Gemini 3 Pro model release. It is positioned not as a successor to Project IDX or a thin layer on top of an existing IDE, but as a new top-level surface in which **agents — not files — are the unit of work** ([Build with Google Antigravity](https://developers.googleblog.com/build-with-google-antigravity-our-new-agentic-development-platform/), [Introducing Google Antigravity](https://antigravity.google/blog/introducing-google-antigravity)). Where Claude Code, GitHub Copilot's agent mode, and Cursor still center the developer on an editor pane and treat the agent as a sidekick that occasionally drives, Antigravity inverts the relationship: the developer issues a high-level intent, the platform spawns one or more parallel agents that operate across the editor, the terminal, and a controlled browser, and the human reviews a stream of persisted **artifacts** that prove what each agent did.
moderate confidence
Azure Foundry Hosted Agents
Microsoft Foundry Hosted Agents is a compute platform purpose-built for running AI agents at enterprise scale. Announced in public preview in April 2026 as part of the Microsoft Foundry platform, it provides every agent session with its own dedicated, hypervisor-isolated sandbox — not process isolation or a code-execution-only container, but full VM-level isolation — with a persistent file system that survives scale-to-zero events. The central insight driving the product is that traditional web-service infrastructure (containers, web apps, serverless functions) was designed to share one instance across many concurrent users, which is incompatible with how agents work: agents write files, execute arbitrary code, hold sensitive credentials, and accumulate session-specific state that cannot be safely shared.
moderate
Claude for Microsoft 365
## What it is
moderate
Claude Managed Agents
## What they are
high
Claude agents
By May 2026 Anthropic's "Claude agents" are no longer a single product but a layered family of tools and primitives that all share one underlying engine. At the bottom is the **Claude Agent SDK** — a runtime, formerly named the Claude Code SDK, that drives an autonomous tool-use loop. Wrapped around it is **Claude Code**, the user-facing agentic coding product that ships as a terminal CLI, a VS Code and JetBrains extension, a desktop app, and a hosted browser/mobile experience at `claude.ai/code`. Inside both surfaces sit four orthogonal extension mechanisms that together decide what the agent can do and when it should do it — **sub-agents**, **Skills**, **hooks**, and **MCP servers** — and an optional fifth, **plugins**, that bundles the other four into shareable packages.
moderate
Codex agents
Codex is OpenAI's umbrella brand for autonomous and semi-autonomous coding agents. Where ChatGPT itself is a generalist chat product, "Codex" denotes a specific agentic stack: a terminal CLI, an IDE sidebar, a hosted cloud agent, and a set of mobile / chat triggers, all sharing the same backend reasoning model, the same on-disk configuration, the same project-level instructions file (`AGENTS.md`), and the same registry of Model Context Protocol (MCP) tools. The product line was relaunched on **May 16, 2025** as a research preview of a cloud software-engineering agent powered by a model called **codex-1**, and over the following year it grew into the multi-surface platform documented here. The information on this page reflects state observable in May 2026; the volatile sections (model lineup, pricing) are explicitly marked.
moderate confidence
Context Architecture: Runtime Retrieval for the Agent Era
## Why RAG Hits Its Limits with Agents
high
copilot cowork
Microsoft 365 Copilot Cowork is Microsoft's branded entry point for long-running, multi-step agentic work inside the Microsoft 365 surface. It launched on May 5, 2026 via the Copilot Frontier early-access program and is presented as a way to "delegate work" to an agent that grounds itself in the user's emails, meetings, messages, files, and data ([Microsoft, 2026-05-05](https://techcommunity.microsoft.com/blog/microsoft365copilotblog/cowork-in-progress/4511672)).
75
GitHub Copilot Spaces: Persistent Context Beyond the Repository
## What Are Copilot Spaces?
moderate
Databricks MCP Marketplace
The **MCP Marketplace** is Databricks' open marketplace for discovering, connecting to, and governing **Model Context Protocol (MCP) servers** alongside AI-ready datasets, announced live in a Platform/Partners blog post on **8 May 2026** by Roman Ostrovski, Harish Gaur, and Antoine Amend. Its purpose is to give agentic applications a way to plug into **real-time external intelligence** — not just static internal data — through a governed, auditable, production-ready channel.
high
Desired State Configuration v3.2.0
## What it is
78
Gemini 3.5: Frontier Intelligence with Action
## Overview
high
GitHub Copilot agents
GitHub Copilot started life in 2021 as an inline autocomplete plugin. Five years later, the brand covers something much wider: a family of agent surfaces that plan multi-step work, run terminal commands, open pull requests on their own, build entire prototype apps from a prompt, and call out to third-party tools through the Model Context Protocol. By May 2026, "Copilot" is not a single product but a layered stack of agentic capabilities that share models, billing, and policy plumbing while differing in *where* they run, *how synchronous* they are, and *how much autonomy* the user grants them.
high
GitHub Copilot Billing — Move to AI Credits (June 2026)
## What it is
high
GitHub Copilot Token-Reduction Tactics
Working methods to cut Copilot token consumption without giving up the tool. All tactics here are configurable by a single developer with normal repo write access — they do **not** require Business / Enterprise admin rights unless explicitly flagged.
moderate confidence
GPT-5.5
OpenAI announced GPT-5.5 on April 23, 2026, positioning it as a fundamental redesign of how AI interacts with operating systems and professional software. GPT-5.5 retakes the benchmark lead from Anthropic's Claude Opus 4.7 (released the previous week) across the majority of publicly available evaluations, with a particular focus on agentic computer use, coding, and scientific research ([VentureBeat, 2026-04-23](https://venturebeat.com/technology/openais-gpt-5-5-is-here-and-its-no-potato-narrowly-beats-anthropics-claude-mythos-preview-on-terminal-bench-2-0)).
moderate
Graphify
## What it is
high
GraphRAG
GraphRAG is a family of retrieval-augmented generation techniques that put a knowledge graph in front of the language model rather than a flat vector index. The name is most closely associated with the [Microsoft Research project of the same name](https://www.microsoft.com/en-us/research/project/graphrag/), introduced in early 2024 and published in the arXiv paper *From Local to Global: A Graph RAG Approach to Query-Focused Summarization* ([Edge et al., 2024](https://arxiv.org/abs/2404.16130)). But “graph-RAG” is now an architectural pattern in its own right, with implementations from Neo4j, LlamaIndex, LangChain, FalkorDB, and others, and a research line that already includes second-generation variants such as DRIFT search and LazyGraphRAG. This page covers the original Microsoft system, its mechanics, why graphs help RAG in the first place, and the 2026 state of the wider ecosystem.
moderate confidence
HTML as AI Output Format
In May 2026 an Anthropic engineering lead who works on Claude Code published a case that HTML — not Markdown — should be the default output format requested from AI agents for most non-trivial work. The argument went viral on X (10.1 million views, 14,200 likes, 28,300 bookmarks) and reached the #1 spot on Hacker News on May 9, triggering widespread discussion across developer communities ([Thariq Shihipar via YouMind, 2026-05-08](https://thariqs.github.io/html-effectiveness/); [Simon Willison, 2026-05-08](https://simonwillison.net/2026/May/8/unreasonable-effectiveness-of-html/); [noqta.tn, 2026-05-10](https://noqta.tn/en/news/anthropic-thariq-html-over-markdown-ai-outputs-2026)).
moderate
LLM Wiki — Karpathy's knowledge-compilation pattern
## What it is
high
m365 content governance 3rs
Microsoft's "3Rs" framework — Readiness, Relevance, Resiliency — is the brand the M365 product group has put around content governance in the Copilot and agents era. The unifying argument is that as AI is embedded in everyday work, governance stops being a back-office function and becomes a precondition for being able to turn AI on at all. The April 22, 2026 release post bundles roughly two dozen distinct capability updates under this framing, spanning SharePoint Advanced Management (SAM), the new SharePoint Admin Agent, Microsoft 365 Archive, and Microsoft 365 Backup ([Microsoft, 2026-04-22](https://techcommunity.microsoft.com/blog/microsoft365copilotblog/microsoft-365-copilot-readiness-and-resiliency-with-sharepoint-and-m365-backupar/4513048)).
moderate confidence
MCP Agent Governance Toolkit (.NET)
The Agent Governance Toolkit (AGT) is an MIT-licensed .NET library that adds a policy enforcement layer to Model Context Protocol (MCP) tool calls inside AI agent systems. Where the MCP specification defines how agents connect to tools, AGT governs what those tools are allowed to do once connected — inspecting inputs and outputs, scanning tool definitions for injection risks, enforcing YAML-based access policies, and emitting OpenTelemetry audit events for every tool call evaluation ([Microsoft .NET Blog, 2026-04-29](https://devblogs.microsoft.com/dotnet/governing-mcp-tool-calls-in-dotnet-with-the-agent-governance-toolkit/)).
moderate
Multi-Agent Systems
This is a hub page on multi-agent AI systems — when to use them, the open protocols that hold them together, and the canonical four-agent reference architecture (LangGraph + MCP + A2A + a local LLM) documented in Sandeep Bharadwaj Mannapur's April 2026 FreeCodeCamp book *How to Build a Multi-Agent AI System with LangGraph, MCP, and A2A*. The architectural reasoning here is the substantive content; the specific tutorial code is just one concrete instantiation of it.
moderate
Pinecone Nexus & KnowQL
## What it is, in one paragraph
high
Power Platform — Custom Tools and Rich UI for App-Based Conversations
## What it is
high
sharepoint ai skills
AI Skills in SharePoint entered Public Preview on April 21, 2026 alongside two adjacent capabilities — site-level *memory* (context) and AI-driven *content generation*. Together they are Microsoft's answer to the recurring feedback from earlier AI-in-SharePoint preview customers: that AI needed to be "more aware of how their organizations operate — the norms, processes, and content preferences that are essential for how work gets done" ([Microsoft, 2026-04-21](https://techcommunity.microsoft.com/blog/spblog/ai-skills-are-now-in-public-preview-teaching-ai-in-sharepoint-what-to-know-and-h/4512532)).
moderate confidence
Unified Agentic Memory Across Harnesses (via Hooks)
This page captures an architectural pattern proposed by **Tomaz Bratanic (Neo4j)** in a *Towards Data Science* post from May 2026: how to give multiple coding-agent "harnesses" — Claude Code and OpenAI Codex CLI — a single shared, portable memory layer by using each harness's **lifecycle hooks** as the integration mechanism, with **Neo4j** as the persistent graph store. The same pattern generalises beyond coding agents.
78
The VS Code Agent Harness: How the Coding Loop Works
## What Is the Coding Harness?
88
Visual Studio Code AI Features (1.110–1.121)
Visual Studio Code underwent a significant architectural shift in early 2026, evolving from a code editor with AI chat features into a platform built explicitly for agentic development. The period from version 1.110 (February 2026) through 1.120 (May 2026) — spanning roughly eleven weekly stable releases — represents the most concentrated wave of AI-centric changes in VS Code's history. GitHub Copilot shifted from an optional extension to a built-in capability, agents gained persistent memory and full session debugging, a dedicated multi-project Agents window graduated from Insiders to Stable, and the platform adopted a philosophy of developer-controlled autonomy from manual approval all the way to fully autonomous Autopilot mode.
high
Windows 365 for Agents
Windows 365 for Agents is a Microsoft Cloud PC offering, purpose-built to host autonomous AI agents on dedicated, IT-managed Windows (and Linux) virtual desktops. It first launched in private preview at **Ignite 2025 (November 18, 2025)** and entered **public preview (US only)** on **May 1, 2026**, announced on the Windows IT Pro blog. The product is positioned as the *runtime layer* of Microsoft's agentic stack: where Microsoft 365 Copilot brings AI into apps and **Microsoft Agent 365** (generally available as of the May 2026 announcement) acts as the control plane, Windows 365 for Agents is the trusted execution environment in which agents actually do work.
82
Work IQ API
Work IQ is the intelligence layer that grounds Microsoft 365 Copilot responses in an organisation's actual work data. In Microsoft's own description it is the "brain behind Copilot" — a three-layer stack of data, context, and skills/tools that runs underneath every Copilot response and agentic workflow ([Microsoft, Mar 2026](https://techcommunity.microsoft.com/blog/microsoft365copilotblog/a-closer-look-at-work-iq/4499789)). The Work IQ API reached public preview on April 30, 2026, exposing that intelligence layer directly to developers so that custom agents and applications can draw on Copilot's permission-aware grounding without wiring raw M365 data sources themselves ([Microsoft Copilot Studio Blog, 2026-04-30](https://techcommunity.microsoft.com/blog/copilot-studio-blog/work-iq-api-public-preview-build-copilot-powered-agents-with-a2a/4516286)).
synthesised
Agentic OS — context, memory, orchestration as a stack
A strategic argument that standalone CLI agents like Claude Code are an "evolutionary dead end" unless they sit inside a broader Agentic Operating System that provides shared context, persistent memory, and multi-agent orchestration. The label is community shorthand; the underlying three-layer pattern is real and converging across major labs.
synthesised
Agent Skills vs MCP — different layers, complementary by design
A practitioner walk-through (Shaw Talebi, narrated as a Dutch-language summary) that resolves the common misread of Anthropic's Skills launch as a competitor to MCP. The video lands on the now-canonical answer: Skills supply the *what to do*, MCP supplies the *how to connect*.
synthesised
Microsoft Copilot Cowork — the agentic workspace inside M365
Copilot Cowork is Microsoft 365 Copilot's longer-running, action-taking mode: rather than describing what could be done, it does the work — sending emails, scheduling meetings, creating documents, posting in Teams — with the user approving each consequential step. The video is a third-party demo; the official Microsoft Learn overview is the authoritative reference.
synthesised
VS Code Agent Skills — the complete guide (Burke Holland)
A Microsoft DevRel-led walkthrough of the Agent Skills format and how it landed in VS Code's GitHub Copilot extension. The video doubles as an authoring tutorial and a comparison guide against adjacent customisation primitives.