Agentic usage with TSANet Connect

Shipped

Agent skills: the part that already exists

An agent skill is a folder with a SKILL.md in it — name and description in frontmatter, instructions underneath, optionally reference files and scripts alongside. An assistant reads the description to decide when the skill applies and loads the full instructions only then. Anthropic released the format as an open standard on 18 December 2025; the specification lives at agentskills.io, and the site lists Claude, OpenAI Codex, Gemini CLI, GitHub Copilot, Cursor and VS Code among the tools that read it. documented That is why the connector pages can say "nothing in it is Claude-specific": the same folder loads into any of them.

TSANet ships four member-facing skills (a fifth, internal one covers the adapters TSANet builds itself and is not part of any kit). Each was distilled from the production implementation it sits next to, and the Zendesk, Fin / Intercom, SDK and Gateway pages on this hub were verified against their skills' repositories. Salesforce and Dynamics have no skill yet.

Loading a skill, whichever assistant you use

git clone --depth 1 https://github.com/tsanetgit/Connect_SDK.git
cp -R Connect_SDK/skills/connect-sdk-deploy  your-project/.claude/skills/

This site

The hub assistant

A chat assistant for developers working from these pages: ask it which connector fits a platform, what a status transition needs, why an auth call fails with a given code, and it answers from the hub, the API reference and the public tsanetgit repositories, with a link to the page it drew on. It is designed as a Microsoft Copilot Studio agent, running on Claude, embedded in every page of this site as an iframe — the site half is shipped, the agent half is not yet created.

Two halves. The site half is shipped: assets/assistant.js renders a launcher in the bottom-right corner and a panel that loads the agent on first open — but only once an embed URL is configured, so no launcher appears for an agent that does not exist. shipped The agent half is not yet created. It has to be built in TSANet's Microsoft tenant by someone with Copilot Studio maker rights, and the recipe below is exactly that job. needs probe

Why Copilot Studio, for a static site

This hub is plain HTML on GitHub Pages with no server. An assistant built directly on a provider's API — Anthropic's, OpenAI's, Google's — needs a hosted backend to hold the API key and run the retrieval; Copilot Studio hosts all of that and hands back a one-line iframe. It also lets TSANet pick the model in a dropdown, including Claude Sonnet 4.6, and change it later without touching this site. documented The trade is that the agent's cost and governance live in TSANet's Microsoft tenant, and the knowledge retrieval runs on Bing's index of these pages rather than on the pages themselves.

Building the agent

1. Admin switches first. Claude appears in the model picker only after the three switches below are on: Anthropic enabled as a subprocessor in the Microsoft 365 admin center (with the maker in the allowed users or groups), external models on for the environment in the Power Platform admin center, and data movement across regions if the tenant is outside the United States. Skip Claude, and the default GPT-4.1 needs none of them. documented
2. Create the agent on the standard harness and pick the model in the Model section of its Overview page. Claude Sonnet 4.6 is the general-category Claude that is GA there; Sonnet 5 needs the GitHub Copilot harness, which does not offer the custom-website embed this site uses. documented
3. Instructions. Something close to this, adjusted as testing shows:

   You help software developers integrate with TSANet Connect. Answer only from
   your knowledge sources: the TSANet Developer Hub, the Connect API reference on
   GitBook, and the public tsanetgit repositories. Link the page you used.
   If the sources do not cover a question, say so and point to the hub's
   Community page; do not invent endpoints, field names, status codes or limits.
   Never ask for or accept credentials, tokens or secrets; credential and
   membership matters go to membership@tsanet.org.
   Prefer the hub's wording where it flags a contradiction between sources, and
   repeat the contradiction rather than resolving it.

4. Knowledge: public websites. Add https://tsanetgit.github.io/tsanet-developer-hub, https://tsanet.gitbook.io/connect and https://github.com/tsanetgit. Public-website knowledge is retrieved through Grounding with Bing Search; the URL may be at most two path levels deep (all three are), the site must be public and indexed by Bing, and a generatively orchestrated agent may hold up to 25 such sources. documented Whether Bing has indexed this hub is unverified; if it has not, upload the pages as file knowledge instead until it has. needs probe
5. Security → Authentication → No authentication. This is what makes the custom-website embed available, and Microsoft's own caution applies verbatim: it "allows anyone who has the link to chat and interact with your bot or agent". That is the intended audience here, but it means the agent's Copilot Credits are spent by anyone who finds it. A Power Platform data policy that requires authentication removes this option entirely. documented
6. Publish, then Channels → Custom website and copy the iframe. Its src has the shape https://copilotstudio.microsoft.com/environments/<env>/bots/<bot>/webchat?__version__=2 and the embed code is shown only while authentication is off. documented Microsoft's current Learn page shows the longer alternative — Bot Framework Web Chat with the token endpoint from Channels → Email — which this site does not need.
7. Wire it into the site. Paste the src into embedUrl in assets/assistant.js, bump the ?v= query on the script tag (the same asset-cache rule as the stylesheet), commit and push. The launcher appears on every page on the next deploy; the script refuses any URL that is not on copilotstudio.microsoft.com.
8. Test from the pages. Ask the questions the hub answers well — the scope that looks right and fails, why responded gates the SLA, why a note has two fields — and check the link it cites. Then ask something the hub does not cover and confirm it says so.

Provider-agnostic

Any provider's agent: what it needs from Connect

Strip away the branding and every agent framework needs the same two things to act on TSANet Connect: a tool definition it can read — an OpenAPI document or a JSON schema per operation — and an authentication path that produces the app-only token Connect expects. The first is easy everywhere. The second is where the frameworks split, because the Connect API's server-to-server path is OAuth 2.0 client credentials: no user signs in, and the service principal is one TSANet provisions and maps to the member company.

The pattern in that last column is the whole finding. Hosted, no-code agent builders assume a signed-in human — Copilot Studio's wizard and custom GPT Actions both offer OAuth only as a redirect-and-consent flow — so they cannot use the client-credentials path as documented — the legacy POST /v1/login JWT remains available, at the cost of something in your system holding and refreshing the token. Code-first frameworks can, because the code that runs the tool is yours and can do the client-credentials exchange itself. The reusable piece, then, is a small service that holds a member's Connect credential and exposes cases, notes, forms and file forwarding as tools. Today the shape most of these frameworks agree on for that service is an MCP server over Streamable HTTP: Copilot Studio and the Claude API both attach one directly documented; the others advertise MCP support that this page has not verified. needs probe Nobody has written that server for Connect. When somebody does, it is the same translation-service work as Working with the API describes, hosted somewhere reachable — gateway-class if TSANet operates it, member-hosted otherwise, which is the open question in the table below.

Design

What no model changes

Everything on the API page still applies when the caller is a language model, and a few of those rules become sharper rather than softer.

• Accepting is a commitment, so the agent must not do it alone. Approval stops the SLA clock and tells a partner an engineer is engaged. The shipped connectors disagree about automated acceptance, and the acceptance topic works through when a rule asserts something true. An agent that can call the approval endpoint on its own judgement is the "accept everything" posture with extra steps; give it the read tools and the note tool, and route acceptance to a person, as the Fin / Intercom connector does.
• The token is the primary key, and the model will not remember it for you. Whichever tool creates a case, the token has to be persisted by something that is not the conversation. There is no idempotency key on create, so a retried tool call is a duplicate case.
• Forms are fetched fresh, because the required documentId arrives with the fetch and partners change their forms. Do not let an agent cache a partner's process form as knowledge.
• Two note fields are not two audiences. summary and description are Subject and Details — both reach the partner, rendered as two sections. Internal-versus-partner visibility is a connector-platform concept (Zendesk models it), not an API one. A tool that "adds a note" is underspecified until it fills both fields.
• Errors look like outages unless you ask otherwise. By default most Connect rejections arrive as HTTP 500; the Accept header that switches on RFC 7807 bodies has to be sent by whatever executes the tool, or the model will read every validation failure as an outage and retry it.

Platform assessment

Microsoft Copilot Studio, with Claude

One hosted way to build an agent over Connect, assessed in detail because it is the platform behind the hub assistant and the one a Dynamics member already has an environment for. Everything here is documented, not observed: read on Microsoft Learn (dates in the sources) and in the source of the shipped Dynamics solution. No Copilot Studio agent has been built against Connect. Assessed 2026-09-13.

Which Claude it offers, and where it runs

A maker picks the agent's primary model from a dropdown on the agent's Overview page, and picks again, separately, for each prompt tool. Since 7 January 2026 Anthropic operates as a Microsoft subprocessor for these models, so Microsoft's Product Terms and Data Protection Addendum apply rather than Anthropic's commercial terms — except for models Microsoft labels "with Data Retention", which stay off by default and need a separate opt-in. documented

• Two Learn pages disagree about Claude's status in prompt tools. The March 2026 release notes say Claude Sonnet and Opus are "generally available globally (excluding GCC)". The prompt-builder settings page, updated in August 2026, says "Anthropic Claude models are at the experimental stage, even though they don't display a tag" — and experimental models are subject to preview terms and not recommended for production. A member deciding whether an agent is production-grade should ask Microsoft which is current. documented
• The same two pages disagree about hosting. The subprocessor page says Microsoft's terms and DPA govern Anthropic models; the prompt-builder page still says "Anthropic models are hosted outside Microsoft and are subject to Anthropic terms and data handling". The subprocessor page is dated later and is the admin-facing one; the hub reports both. documented
• Content moderation is Microsoft's slider, and it is unavailable for Claude. The per-prompt moderation level applies only to Microsoft-managed models. Temperature still applies. documented
• Revoking Anthropic after agents are built does not stop them cleanly. Agents switch to "another suitable internal model" if one exists and error out if not, until a maker changes the model by hand. documented

Three switches, in two admin centres, before a maker sees Claude

• Access is per user, not just per tenant. The subprocessor setting takes a user or security-group assignment, enforced across Copilot Studio: a user outside it cannot use an agent that runs on Claude. For the hub assistant that constraint falls on the maker, not the anonymous visitor. documented
• Preview models and external models are separate gates. Sonnet 4.6 is GA and external, so it needs only the external-model gate. documented
• Compliance exclusions are explicit. Anthropic models in Copilot Studio are outside the EU Data Boundary; FedRAMP is not achieved; PCI DSS does not apply; the government-cloud table lists no Claude model. A member in any of those positions should stop at this row. documented

Three routes from an agent to the Connect API

Copilot Studio gives an agent tools; what a TSANet agent needs is a tool that reaches Connect. There are three shapes, and the platform's own rules decide which one a member can use.

The harness decides the model and the tools together

Copilot Studio runs agents on one of two harnesses, and since June 2026 the choice is real. The Claude model and the Connect tooling are not chosen independently: each harness admits a different Claude and a different set of tool types. documented

• Sonnet 5 costs you the REST API tool and the iframe embed. Both are documented for the standard harness only. An agent that wants Sonnet 5 is on the newer harness, where the documented ways to reach an external API are an MCP server or a workflow. documented
• A conversational turn is many model calls. The standard harness does not enforce a token limit across the planner, retrieval and tool-summary calls one user message triggers; the 200K context applies per call. documented
• Every new agent gets a Microsoft Entra Agent ID from July 2026, and the environment-level opt-out is gone. It is not a Connect identity and does not replace the service principal TSANet provisions. documented

Open questions

What has to be answered, and what each one blocks

Verified against

Sources, with the dates that were read

Vendor documentation changes monthly; the dates are each page's own date stamps at the time of reading. Where a row here contradicts the live page, the live page wins and this one needs revisiting.