Copilot Harness and SDK Map

1The five things people conflate

The one sentence that resolves the confusion. A/B/C give you the agent loop inside your process under GitHub identity and billing; D gives you a Microsoft-managed SaaS agent, named after the same harness, under Entra identity, Power Platform governance, and Copilot Credits. E is how code talks to D, and today it is only officially supported for the standard harness.

What is documented versus inferred about A and D. No Microsoft page states that the Copilot Studio harness is the Copilot CLI runtime. The evidence that they share a lineage is the name, the cli-copilot authoring mode that pac copilot init uses for it, the CLICopilotRecognizer / CLIAgentRecognizer recognizer kinds in settings.mcs.yml, and Microsoft's own plugin calling these "CLI (agentic-loop) agents". Treat "same runtime" as an inference; treat "different surface, identity, billing, and API" as documented. The independent verification pass run for this document reached the same conclusion ("no verified source states they are the same runtime").

Microsoft's own definition of a harness: "a runtime that exists between the two [your design and the model]: it determines when to call the model, what components to send it, interprets what comes back, and calls the right tools." (Learn: Harnesses in Copilot Studio)

The blog's split of responsibilities between B and C: "Copilot owns the agent loop (model calls, tool invocation, planning, and session state) while Agent Framework gives you a consistent surface for instructions, tools, streaming, middleware, observability, and human-in-the-loop approval." (devblogs, 4 Aug 2026)

2Capability map: the same idea in each surface

Read across a row to find the equivalent concept. "—" means the surface does not expose it.

Sources for this table: SDK docs (getting-started, features, hooks, auth, setup); Agent Framework provider page (Learn, updated 25 Aug 2026); Copilot Studio agents-experience pages (overview, tools, skills, memory, connected agents, model).

3GitHub Copilot SDK (B): what it is and how to drive it

3.1 Packages and prerequisites

• License MIT. "All SDKs communicate with the Copilot CLI server via JSON-RPC." SDK and CLI negotiate protocol versions at startup; SDK protocol range v2–v3, CLI v3 fully supported, v2 via adapters (compatibility).
• Who can use it: "all existing GitHub Copilot subscribers, including Copilot Free for personal use, and non-Copilot users via BYOK" (GA changelog, 2 Jun 2026).
• Security note from Microsoft Learn: "it is recommended to run agents with shell or file permissions in a containerized environment (Docker/Dev Container)."

3.2 First program (TypeScript)

import { CopilotClient } from "@github/copilot-sdk";

const client = new CopilotClient();
const session = await client.createSession({ model: "auto" });
const response = await session.sendAndWait({ prompt: "What is 2 + 2?" });
console.log(response?.data.content);

await client.stop();

.NET equivalent uses CopilotClient, CreateSessionAsync(new SessionConfig { Model = "auto", OnPermissionRequest = PermissionHandler.ApproveAll }), and SendAndWaitAsync(new MessageOptions { Prompt = ... }). Python uses CopilotClient(), await client.start(), create_session(on_permission_request=PermissionHandler.approve_all, model="auto"), send_and_wait(...). (getting-started)

3.3 Client options (process-level)

Environment variables the runtime reads: COPILOT_GITHUB_TOKEN → GH_TOKEN → GITHUB_TOKEN (in that order after an explicit token and a direct GITHUB_COPILOT_API_TOKEN + COPILOT_API_URL), GITHUB_COPILOT_INTEGRATION_ID (attribution header Copilot-Integration-Id), COPILOT_PLUGIN_DIR_ONLY=true (deterministic plugin set). (authenticate, cloud-sessions, plugin-directories)

3.4 Session options (per conversation)

Sources: session-persistence, multi-tenancy, custom-agents, skills, mcp, session-limits, cloud-sessions, context-management.

3.5 The agent loop and its events

"Each iteration of this loop is exactly one LLM API call, visible as one assistant.turn_start / assistant.turn_end pair." The loop ends with session.idle (always emitted, ephemeral, returned by sendAndWait). session.task_complete is optional, persisted, and requires the model to signal it. (agent-loop)

Verified in the 1.0.13 source (dist/session.js): sendAndWait treats session.idle as the end of the request only when event.data.mode !== "autopilot"; an idle emitted while the agent is in autopilot mode is a pause between steps, not completion. IdleData is { aborted?: boolean, mode?: SessionMode }. Code that ends a turn on any idle will truncate autopilot runs.

Event envelope: id, timestamp, parentId, type, data, optional agentId, ephemeral. Ephemeral events are not replayed on resume; persisted ones are recoverable via getMessages.

Subscribe: TS session.on("assistant.message_delta", e => ...); Python session.on(handler) and filter on event.type; .NET session.On<AssistantMessageDeltaEvent>(...).

3.6 Hooks

Nine hooks, registered via hooks: in session config (hooks-overview): onPreToolUse, onPostToolUse, onPostToolUseFailure, onUserPromptSubmitted, onUserPromptTransformed, onSessionStart, onSessionEnd, onErrorOccurred, onAgentStop.

onPreToolUse input: timestamp, cwd, toolName, toolArgs (+ invocation sessionId). Output: permissionDecision ("allow" | "deny" | "ask"), permissionDecisionReason, modifiedArgs, additionalContext, suppressOutput. Return null to allow unchanged. (pre-tool-use)

Permission requests observed live here (SDK 1.0.13 on CLI 1.0.84, 6 Sep 2026, via sdk/examples/copilot-sdk.mjs): a custom tool defined with defineTool arrives at onPermissionRequest as { kind: "custom-tool", toolCallId, toolName, toolDescription, ... }. Returning { kind: "approve-once" } (the value the SDK's own approveAll helper returns) let the tool run (tool.execution_complete with success: true, answer "Order 42 is currently shipped."); returning { kind: "denied-interactively-by-user" } produced success: false and the model explained the refusal. approveAll throws when managed settings are enabled and returns { kind: "no-result" } when a request is flagged managedApprovalRequired. The full PermissionDecision union in the 1.0.13 typings: ApproveOnce, ApproveForSession, ApproveForLocation, ApprovePermanently, Reject, UserNotAvailable, Approved, ApprovedForSession, ApprovedForLocation, Cancelled, DeniedByRules, DeniedNoApprovalRuleAndCouldNotRequestFromUser, DeniedInteractivelyByUser, DeniedByContentExclusionPolicy, DeniedByPermissionRequestHook (kind strings are the kebab-case forms, e.g. approve-once, approve-for-session). Set skipPermission: true on a tool definition to bypass the prompt for that tool. model: "auto" resolved to mai-code-1.1-flash in these runs.

3.7 Authentication and licensing

Priority order (authenticate):

1. Explicit gitHubToken on client or session.
2. Direct API token: GITHUB_COPILOT_API_TOKEN with COPILOT_API_URL.
3. COPILOT_GITHUB_TOKEN → GH_TOKEN → GITHUB_TOKEN.
4. Stored credentials from copilot CLI login.
5. gh auth credentials.

Troubleshooting S2S: 401 = org does not support GitHub App auth; 403 Resource not accessible = token in wrong env var; 403 from the Copilot API = missing repository_ids/permissions or installation lacks "All repositories".

3.8 Bring your own key

provider shape (byok):

Under BYOK: model availability, rate limits, and usage tracking come from your provider; "Premium requests do not count against Copilot premium request quotas." Keys are not persisted with the session.

Docs disagree on Entra. The repository README says "BYOK uses key-based authentication only. Microsoft Entra ID (Azure AD), managed identities, and third-party identity providers are not supported," while docs/setup/azure-managed-identity.md documents bearerTokenProvider with DefaultAzureCredential for Foundry. Treat managed identity as available through the bearer-token path and test it before depending on it.

3.9 Running it as a service

• Start the runtime: copilot --headless --port 4321 (binds loopback; add --host 0.0.0.0 for other machines). Set COPILOT_GITHUB_TOKEN before starting. Add --session-idle-timeout <seconds>. There is "no built-in auth between SDK and CLI", so secure the TCP path. (backend-services)
• Multi-user: one runtime, mode: "empty", unique session ids per user, gitHubToken per session, explicit availableTools. Session state lives at COPILOT_HOME/session-state/{sessionId}; put it on shared storage (Azure Files, NFS, PVC) if any server must resume any session. No built-in session locking or load balancing; CLI is single-process. (multi-tenancy, scaling)
• Health: client.getStatus() (also returns protocolVersion). Cleanup: client.listSessions() + client.deleteSession(id).
• Usage: assistant.usage per model call (inputTokens, outputTokens, model, cost multiplier); session.usage_info (currentTokens, tokenLimit); session.usage.getMetrics() (totalNanoAiu, totalPremiumRequestCost, modelMetrics); models.list (billing multipliers, token prices); account.getQuota() (quotaSnapshots.premium_interactions, remainingPercentage, resetDate). Nano-AI units divide by 1e9 for credits. (usage-and-billing)
• Plugins: a directory with plugin.json (or .github/plugin.json), optional SKILL.md, hooks.json, .mcp.json, agents/*.md, skills/*/SKILL.md, .lsp.json; loaded with --plugin-dir (ephemeral, precedence over marketplace plugins). (plugin-directories)

4Agent Framework GitHub Copilot provider (C)

4.1 Install and construct

Prerequisites (blog): Copilot CLI installed, active Copilot subscription, authenticated CLI. Python env vars: GITHUB_COPILOT_CLI_PATH, GITHUB_COPILOT_MODEL (e.g. gpt-5, claude-sonnet-4), GITHUB_COPILOT_TIMEOUT (default 60 s), GITHUB_COPILOT_LOG_LEVEL, GITHUB_COPILOT_BASE_DIRECTORY (default ~/.copilot). .NET: CopilotClientOptions { CliPath, LogLevel, BaseDirectory, WorkingDirectory }. (Learn provider page)

4.2 Canonical .NET sample (from the blog)

using GitHub.Copilot;
using GitHub.Copilot.Rpc;
using Microsoft.Agents.AI;

await using CopilotClient copilotClient = new();
await copilotClient.StartAsync();

SessionConfig sessionConfig = new()
{
    OnPermissionRequest = (request, invocation) =>
        Task.FromResult(PermissionDecision.ApproveOnce()),
};

AIAgent agent = copilotClient.AsAIAgent(sessionConfig, ownsClient: true);
AgentResponse response = await agent.RunAsync("Summarize what this project does.");
Console.WriteLine(response);

4.3 Semantics you must know

• Default is deny-all. "By default, the agent cannot execute shell commands, read/write files, or fetch URLs" until you supply OnPermissionRequest / on_permission_request. Decision types: .NET PermissionDecision.ApproveOnce(), .Reject(feedback?), .UserNotAvailable() (the default when no handler is set); Python PermissionHandler.approve_all, PermissionDecisionDeniedInteractivelyByUser, PermissionDecisionUserNotAvailable; Go PermissionDecisionApproveOnce, PermissionDecisionReject, PermissionDecisionDeniedInteractivelyByUser.
• Tool approval runs in the SDK hook, not the framework round-trip. Wrapping a tool in ApprovalRequiredAIFunction (.NET) or @tool(approval_mode="always_require") (Python) makes the provider install a default OnPreToolUse hook that returns "ask" and routes to your permission handler. If you register your own pre-tool hook, that default is not installed and you own enforcement; the agent logs a warning naming each approval-required tool.
• default_options is a passthrough to the SDK's create_session: reasoning_effort, context_tier, enable_citations, provider, skill_directories, mcp_servers, etc. Unknown keys raise TypeError.
• Sessions: .NET AgentSession session = await agent.CreateSessionAsync(existingSessionId); Python session.service_session_id round-trips through agent2.get_session(service_session_id=...).
• Context providers (Python): context_providers=[InMemoryHistoryProvider()] run before/after each invocation.
• Observability: Python configure_otel_providers(enable_console_exporters=True); RawGitHubCopilotAgent for the untraced agent.
• Tool support table: Function tools ✅, tool approval ✅, shell/file/URL ✅ (gated), hosted (HTTP) MCP ✅, local (stdio) MCP ✅, code interpreter ❌, file search ❌, web search ❌ ("Not a Copilot CLI capability" / "Not exposed as a hosted tool").
• Samples: .NET samples/02-agents/AgentProviders/github-copilot; Python samples/02-agents/providers/github_copilot in microsoft/agent-framework. Release cadence at time of writing: dotnet-1.20.0 (31 Aug 2026), python-1.17.0 (3 Sep 2026).

4.4 What Agent Framework adds versus using the SDK directly

Per the SDK's own integration page (microsoft-agent-framework.md): sequential/concurrent orchestration, multi-provider composition (Copilot beside Azure OpenAI, Anthropic, Copilot Studio), A2A protocol, one AIAgent abstraction. The SDK keeps: tool definition, MCP config, custom agents, infinite sessions, model selection, streaming deltas, permissions. Integration packages exist for .NET and Python only; TypeScript, Go (framework has its own provider), Java, and Rust use the SDK directly.

5Copilot Studio GitHub Copilot harness (D)

5.1 What Microsoft says it is

• "The GitHub Copilot harness is the most capable option, built for reasoning-heavy agents and workflows... it can take a goal, break it into steps, call the right tools across connectors, knowledge, MCP, and connected agents, and adjust when a step fails or a request changes. It natively creates and edits Word, Excel, PowerPoint, and PDF files, supports skills and memory, and runs each task in a secure sandbox governed by Copilot Studio. Agents and workflows on this harness use Copilot Credits." (harnesses-overview)
• Three harnesses: GitHub Copilot harness (reasoning-heavy agents and workflows), standard harness (rule-based agents and agent flows), Copilot chat harness (extend Microsoft 365 Copilot Chat).
• Agents cannot be transferred between harnesses; you choose at creation. The GitHub Copilot harness "uses this enhanced orchestration model for all agents" (not configurable). (agents-experience/overview)
• Timeline: production-ready preview from June 2026; GA announced 3 Aug 2026; usage-based billing in Copilot Credits for all such agents from 1 Sep 2026 (agents created before 3 Aug kept old pricing only until then). (Tech Community announcement, billing overview, 27 Aug 2026)

5.2 Authoring surface

Components and their exact scope:

• Tools = Connectors (Power Platform connectors), MCP servers, Workflows. Added via Build → Tools → Add a tool with Featured / MCP / Connectors / Workflows tabs. (tools-available)
• Skills = SKILL.md (YAML front matter name, description, Markdown instructions) plus optional scripts/templates/reference docs, packaged as ZIP; authored in Studio or uploaded. Same file format the Copilot SDK/CLI uses for skills. (skills-overview)
• Memory (preview): per-user folder in Microsoft-managed storage; capture → store → apply; private to the user; deleted after 28 days of inactivity; disabled in group chats and Teams channels; a memory portal link appears on first interaction per channel. (memory-overview, 27 Jul 2026)
• Connected agents: "you can currently only connect other agents built in Copilot Studio." (authoring-add-other-agents)
• Model: chosen per agent; admins gate preview/experimental models (environment setting plus cross-region data movement) and external providers (Anthropic, Mistral, xAI via Microsoft 365 admin center). (authoring-select-agent-model)

5.3 Billing

"Copilot credits are charged for large language model (LLM) tokens, tools (including knowledge and MCPs), and the harness itself." "Billing starts when you start building": previewing, testing, and generating evaluations all consume credits, unlike the standard harness which bills after publish. Credits are allocated per environment in the Power Platform admin center. (billing-credit-overview)

5.4 Publishing and channels (as of the 3 Aug 2026 page update)

Source: publication-channels-overview. Publishing creates a versioned live copy; DLP policies are evaluated per channel; the built-in Direct Line test channel does not block publishing. (publication-publish-agent, publish overview, 31 Aug 2026)

6Agents as code for the GitHub Copilot harness

The web UI is not the only authoring path. The Power Platform CLI and an experimental Microsoft plugin let you treat a harness agent as files.

6.1 Power Platform CLI (pac copilot)

Installed locally: pac 2.10.1. Subcommands: list, init, status, pack, create, publish, pull, push, clone, delete, quarantine, model, mcp, translation helpers. The plugin requires pac ≥ 2.9.3 (copilot-studio-plugin README).

• pac copilot init with authoring mode cli-copilot creates an empty GitHub-Copilot-harness agent workspace in a target environment (the plugin's init sub-agent uses exactly this; publisher prefix defaults to catmgr).
• pac copilot clone pulls a published agent into a local workspace; push/pull sync; publish publishes.

6.2 Workspace files that identify a harness agent

From the plugin's chat script and skills (chat-with-agent.js, commands/chat.md):

6.3 The experimental plugin (mcs-assistant)

/plugin marketplace add microsoft/copilot-studio-plugin
/plugin install mcs-assistant@copilot-studio-plugin

Works in GitHub Copilot CLI and Claude Code. Provides /migrate (clone a classic agent, describe it, approve a migration plan, convert actions to tools, author the new YAML, push) and /chat (stream a turn against the published harness agent through the /3p route with MSAL device-code sign-in). Sub-agents: architect, describer, init, manage. Disclaimer: "an experimental research project, not an officially supported Microsoft product... not meant for production use."

Related but different: the Copilot Studio VS Code extension (GA) and microsoft/skills-for-copilot-studio cover the standard harness only; the plugin README calls the latter the legacy plugin that "may conflict". (VS Code extension overview, 24 Jul 2026)

7Reaching a harness agent from code: the integration map

7.1 Official position

• Copilot Studio client library: "Currently, you can only use the Copilot Studio client library with Copilot Studio agents created by using the standard harness. Agents using the GitHub Copilot harness aren't yet officially supported." (Learn, 21 Aug 2026)
• Direct Line / native app channel for harness agents: No (§5.4).
• Agent Framework Copilot Studio provider: invokes a published agent and "does not expose Agent Framework tool types... at the client"; capabilities are configured on the Copilot Studio agent. (Learn)
• Microsoft CAT decision guide (updated 2 Aug 2026): the SDK client "requires delegated (user) authentication through Entra ID, with no service principal auth, app-only tokens, or secure anonymous options"; app-only "is on the roadmap". (mcscatblog)

7.2 The experimental route that works today: Agentic Runtime /3p

Microsoft's own plugin talks to "the agenticruntime '3p' (third-party) Direct-to-Engine endpoint, which is how CLI-authored / agentic-loop agents are served." This repository reproduced it in three stacks.

https://{env-host}/copilotstudio/agenticruntime/3p/dataverse-backed/authenticated/bots/{schemaName}?api-version=1

• {env-host} = environment GUID with dashes removed, split into the first 30 and last 2 hex characters: {30}.{2}.environment.api.powerplatform.com for Prod. Other clouds use api.test., api.preprod., api.dev. etc. This repo maps Prod only.
• Token: delegated Entra user token for scope https://api.powerplatform.com/.default, from an app registration holding the Power Platform API delegated permission CopilotStudio.Copilots.Invoke. Public client (SPA or native) with no secret.
• Agent prerequisites: published, Authenticate with Microsoft, and shared with the signed-in user. Unshared → 403.
• The client library appends /conversations and /conversations/{id}; keep api-version=1 pinned. The conversation id also rides the x-ms-conversationid response header.
• Streaming shape differs by client: Node @microsoft/agents-copilotstudio-client delivered cumulative typing snapshots; Agent Framework CopilotStudioAgent.RunStreamingAsync delivered delta fragments in AgentResponseUpdate with the original activity in RawRepresentation (15 fragments consolidated to a 1,164-character answer in the 6 Aug 2026 verification).

7.3 What does not work (and the error you will see)

7.4 Client library connection settings (E)

ConnectionSettings in @microsoft/agents-copilotstudio-client (API reference):

.NET equivalents used in this repo's sidecar: new ConnectionSettings { DirectConnectUrl, Cloud = PowerPlatformCloud.Prod, CopilotAgentType = AgentType.Published }, new CopilotClient(settings, httpClientFactory, logger, httpClientName), new CopilotStudioAgent(client, loggerFactory).

Python Agent Framework: pip install agent-framework-copilotstudio --pre; env COPILOTSTUDIOAGENT__ENVIRONMENTID, __SCHEMANAME, __AGENTAPPID, __TENANTID; CopilotStudioAgent() and agent.run(..., stream=True). .NET quick form: new CopilotStudioChatClient(environmentId, agentIdentifier, credential: new AzureCliCredential()).AsAIAgent(...).

8Decision guide

9Capabilities you can leverage now, and gaps to plan around

Available now (with the surface that provides it)

• Streaming token deltas with 40+ typed events, per-call token usage, and account quota (SDK).
• Custom tools with JSON schemas, approval-gated tools, and nine lifecycle hooks (SDK, Agent Framework).
• Local stdio and remote HTTP/SSE MCP servers with per-server tool allow-lists (SDK, Agent Framework, Copilot Studio).
• Sub-agents with scoped tools, per-agent models, fleet-mode parallelism (SDK).
• Skills as SKILL.md in both worlds; plugin directories bundling skills, agents, hooks, MCP, LSP (SDK).
• Resumable sessions on disk, infinite sessions with compaction, budget caps (SDK).
• Headless multi-tenant hosting, GitHub-hosted cloud sessions with Mission Control URLs (SDK).
• BYOK to OpenAI, Azure OpenAI/Foundry (including Entra bearer tokens), Anthropic, local OpenAI-compatible servers (SDK).
• Org-billed server-to-server identity without user seats (SDK via GitHub App or Actions).
• Maker-authored agents with connectors, MCP, workflows, skills, per-user memory, native Office file editing, model choice including external providers, evaluation and monitoring (Copilot Studio).
• Delegated-token streaming into custom canvases via /3p (verified in this repo, experimental).

Gaps and caveats

• Copilot Studio harness agents: no Direct Line/native app channel, no MCP-client exposure, client library not officially supported, connected agents limited to Copilot Studio agents, memory off in Teams channels, credits consumed during authoring.
• App-only access to harness agents: not available outside a no-auth private preview.
• Agent Framework GitHub Copilot provider: no code interpreter, file search, or hosted web search; Python package pins an older github-copilot-sdk (1.0.2) than the SDK's latest (1.0.13).
• Agent Framework Copilot Studio provider: still preview (1.20.0-preview.*); tools live on the agent, not the client.
• SDK: no built-in session locking or load balancing; BYOK keys not persisted; README and setup docs disagree on Entra for BYOK; installation tokens must go through the env var, not the gitHubToken option.
• SDK 1.0.13 behavior verified here: constructing CopilotClient with mode: "empty" throws unless baseDirectory or sessionFs is set ("Empty mode requires an explicit per-session persistence location"), and createSession in that mode throws unless availableTools is set explicitly (for example ["custom:*"] or new ToolSet().addBuiltIn(BuiltInTools.Isolated)). The multi-tenancy doc shows availableTools but does not state either requirement.
• SDK 1.0.13 typings expose far more event types than the ~40 the streaming-events doc lists (canvas, factory, fusion, MCP OAuth, schedules, plan mode, and more); the documented set in §3.5 is the stable subset to build on.
• Stale vendor pages: the docs.github.com Agent Framework integration page still shows --prerelease and an AIAgentOptions overload that does not exist in source.
• The Copilot Studio harness needs no GitHub Copilot seat; Copilot Credits are the only meter, and dev/trial non-billing ended 1 Sep 2026.

10Version ledger (checked 6 September 2026)

11Sources

GitHub

• https://github.com/github/copilot-sdk (README, MIT, language table)
• https://github.com/github/copilot-sdk/tree/main/docs (getting-started, auth/, setup/, features/, hooks/, integrations/microsoft-agent-framework.md, troubleshooting/compatibility.md)
• https://github.blog/changelog/2026-06-02-copilot-sdk-is-now-generally-available/
• https://github.com/github/copilot-cli/releases
• https://docs.github.com/copilot/how-tos/copilot-sdk

Microsoft Agent Framework

• https://devblogs.microsoft.com/agent-framework/build-production-ready-agents-with-the-github-copilot-harness-and-agent-framework/
• https://learn.microsoft.com/en-us/agent-framework/integrations/by-component/agent-services/github-copilot
• https://learn.microsoft.com/en-us/agent-framework/integrations/by-component/agent-services/copilot-studio
• https://github.com/microsoft/agent-framework/releases

Copilot Studio

• https://learn.microsoft.com/en-us/microsoft-copilot-studio/harnesses-overview
• https://learn.microsoft.com/en-us/microsoft-copilot-studio/agents-experience/overview
• https://learn.microsoft.com/en-us/microsoft-copilot-studio/agents-experience/tools-available
• https://learn.microsoft.com/en-us/microsoft-copilot-studio/agents-experience/skills-overview
• https://learn.microsoft.com/en-us/microsoft-copilot-studio/agents-experience/memory-overview
• https://learn.microsoft.com/en-us/microsoft-copilot-studio/agents-experience/authoring-add-other-agents
• https://learn.microsoft.com/en-us/microsoft-copilot-studio/agents-experience/authoring-select-agent-model
• https://learn.microsoft.com/en-us/microsoft-copilot-studio/agents-experience/billing-credit-overview
• https://learn.microsoft.com/en-us/microsoft-copilot-studio/agents-experience/publication-channels-overview
• https://learn.microsoft.com/en-us/microsoft-copilot-studio/agents-experience/publication-publish-agent
• https://learn.microsoft.com/en-us/microsoft-copilot-studio/agents-experience/publication-fundamentals-publish-channels
• https://learn.microsoft.com/en-us/microsoft-copilot-studio/visual-studio-code-extension-overview
• https://learn.microsoft.com/en-us/microsoft-365/agents-sdk/integrate-with-mcs
• https://learn.microsoft.com/en-us/javascript/api/@microsoft/agents-copilotstudio-client/connectionsettings?view=agents-sdk-js-latest
• https://techcommunity.microsoft.com/blog/copilot-studio-blog/more-powerful-agents-and-workflows-for-autonomous-business-processes-introducing/4542969
• https://microsoft.github.io/mcscatblog/posts/copilot-studio-api-decision-guide/
• https://github.com/microsoft/copilot-studio-plugin (README, commands/chat.md, commands/migrate.md, agents/copilot-studio-init.md, scripts/src/chat-with-agent.js)
• https://github.com/microsoft/skills-for-copilot-studio

This repository

• README.md (live verifications of 5, 6, and 30 Aug 2026), sidecars/AgentFrameworkGhcp/Program.cs, docs/ghcp-harness-bom-and-setup.html, docs/dynamics-sidepane-deployment.html