Make A365 Agent Skill for Claude

Make A365 Agent

> Trigger phrases — any of these will activate this skill:

> - "provision this agent with agent 365"

> - "register this agent"

> - "make this agent findable in the Agent 365 catalog"

> - "Registration setup for this agent"

> - "make this a custom engine agent"

> - "run a365 setup all"

> - "create a365 blueprint for this agent"

> - "set up this agent for observability only"

> What this skill does: Provisions your agent with Agent 365 — creates the Blueprint

> and Entra ID permissions. After setup, always offers observability

> (instrument-observability) and WorkIQ tools (add-workiq-tools) as optional add-ons.

>

> This skill is normally called from a365-setup after CLI verification and Azure

> prerequisites are confirmed. It can also be invoked directly when those steps are done.

> YOUR FIRST ACTION: Load context (Phase 0), then create all todos before running anything.

Phase 0 — Load Context

> Show the user a visible task checklist BEFORE Phase 1 work begins. This skill has no per-phase TaskCreate calls in the body — derive the checklist from the phase headers (## Phase 0 — Load Context, ## Phase 1 — Collect Provisioning Inputs, ## Phase 2 — Register with Agent 365, etc.). Exactly one item in_progress at a time; complete before moving on.

> - Claude Code: call TaskCreate once per phase header (already in allowed-tools); the list renders natively. Use TaskUpdate to flip statuses.

> - VS Code Copilot Chat / GitHub Copilot CLI: allowed-tools is ignored — emit a markdown checklist directly in chat (- [ ] Load context…, - [ ] Collect provisioning inputs…, etc.) and edit items to - [x] as each phase completes.

Check for context passed from a365-setup. If this skill was invoked by a365-setup,

the session already has capabilities, agentStack, programmingLanguage, and

usesTeamsOrCopilot set. Load those values directly.

If invoked directly (no session context):

• Read .a365-workspace-detection.local.json if it exists and detectedAt is within 60 minutes.

Load agentStack, programmingLanguage, and usesTeamsOrCopilot from it.

• If no fresh cache, ask the user in a single message:

What capabilities would you like to enable? (options can be combined)

  1. Register — make the agent findable in the Agent 365 catalog
  2. Observability — end-to-end activity tracing for every message, LLM call,
     and tool use, visible in the Agent 365 portal and Microsoft Defender
  3. WorkIQ — add WorkIQ MCP servers (M365 data: email, calendar, Teams, SharePoint, OneDrive)
     (only show this option when authMode ≠ s2s — WorkIQ requires a user token)
  4. AI Teammate — the agent needs a first-class M365 identity
     (Agentic User with UPN, mailbox, presence). Handled by a different skill.

• If the user selects option 4 (AI Teammate) — stop here and tell them:

> "AI Teammate setup is handled by the make-ai-teammate skill. Run /agent365:a365-setup and select the AI Teammate path, or invoke make-ai-teammate directly."

• Otherwise store the answer as capabilities and continue.

Create all todos for this session:

• Todo 1: Register agent with Agent 365 (a365 setup all)
• Todo 2: Add Observability (optional)
• Todo 3: Add WorkIQ Tools (optional)

Mark Todo 1 in-progress.

Phase 1 — Collect Provisioning Inputs

1.0 — Check for Existing Blueprint

Before asking for any inputs, check whether a blueprint config already exists:

ls a365.config.json a365.generated.config.json 2>/dev/null

If either file exists, read it and extract the blueprint ID — the field name differs by file:

• a365.config.json → read blueprintId
• a365.generated.config.json → read agentBlueprintId

Store whichever is present as existingBlueprintId. Then ask:

I found an existing Agent 365 config in this project.
  • File: {filename found}
  • Blueprint ID: {existingBlueprintId if found, otherwise "not yet set"}

What would you like to do?

  1. Reuse the existing blueprint — I'll skip `a365 setup all` and use this blueprint directly
     (use this if setup already ran successfully and you just want to add capabilities)
  2. Create a fresh blueprint — runs `a365 setup all` and overwrites the existing config
     (use this if you want to start over or the existing config is stale)

Wait for the answer:

• If 1 (reuse): if agentBlueprintId is empty, ask "Please provide your blueprint ID." Store as existingBlueprintId. Set reuseBlueprint = true. Write both values back to .a365-workspace-detection.local.json (merge, preserve all other fields) so the stop-hook validator and follow-on skills can read them. Skip Phase 2 (setup all) entirely — proceed directly to Phase 3.
• If 2 (fresh): set reuseBlueprint = false. Write reuseBlueprint: false to .a365-workspace-detection.local.json. Continue with Phase 1 inputs and Phase 2 as normal.

If no existing config is found: set reuseBlueprint = false and continue.

Ask both questions in a single message:

To provision your agent with Agent 365, I need two things:

  1. Agent Name — short, unique identifier for your tenant (e.g. "contoso-hr-agent" or "FabrikamHelpdesk")
     Rules: letters, numbers, hyphens only. Start with a letter. 3–20 chars (the CLI appends " Blueprint" to derive the Teams manifest `name.short`, capped at 30).
     This derives the Blueprint name. Pass the name exactly as you type it — do NOT normalize case.
     Type "default" to use the name "developer".

  2. Project directory — full path to your agent code, or "current" for this directory.

Store as agent_name and project_dir. If the user replies current, use CWD.

If the user types default, set agent_name = "developer".

1.1 — Determine Messaging Endpoint

Ask the user where their agent is (or will be) hosted:

Where will your agent run?

  1. Azure / Cloud — the agent has (or will have) a public HTTPS endpoint already
  2. Local / Dev Tunnel — the agent runs on localhost and needs a dev tunnel for a public URL
  3. Skip - I just want to set up the Blueprint and Entra ID permissions for now, without registering an endpoint yet

If Cloud (option 1): Ask for the full HTTPS endpoint URL (e.g. https://myagent.azurewebsites.net/api/messages). Store as messagingEndpoint.

If Local / Dev Tunnel (option 2): auto-start the tunnel — do NOT ask the user to paste a URL.

Dev Tunnel Setup (auto-start)

• Verify CLI is installed: devtunnel --version. If it fails, install it and stop until the user confirms:

| OS | Install command |

|----|-----------------|

| Windows | winget install Microsoft.devtunnel |

| macOS | brew install --cask devtunnel |

| Linux | curl -sL https://aka.ms/DevTunnelCliInstall | bash |

• Verify login: devtunnel user show. If it exits non-zero or prints "not logged in", run devtunnel user login (or devtunnel user login --device-code on headless machines), wait for the user to complete sign-in, then retry devtunnel user show.

• Create the tunnel + port (idempotent — treat "already exists" as success). The port is the agent's listening port — 3978 for Node.js / Python, the .NET project's launch port for .NET (default 5000):

   devtunnel create <agent-name>-tunnel --allow-anonymous
   devtunnel port create <agent-name>-tunnel -p <port>

Parse the Tunnel ID from the create output — format is <id>.<cluster> (e.g. abc123xy.usw3).

• Start hosting in the background — this is a long-running process. Run with run_in_background=true so the tunnel keeps running while the skill continues; the user does NOT need a separate terminal:

   devtunnel host <agent-name>-tunnel

• Resolve the public URL deterministically — it is https://<id-without-cluster>-<port>.<cluster>.devtunnels.ms. Example: tunnel ID abc123xy.usw3 + port 3978 → https://abc123xy-3978.usw3.devtunnels.ms. Sanity-check with devtunnel show <agent-name>-tunnel and confirm the printed Access URL matches.

• Store tunnelUrl and set messagingEndpoint = "${tunnelUrl}/api/messages". Tell the user verbatim: *"Dev tunnel started at <URL>. Hosting in the background — leave this session open. Using this endpoint for a365 setup all."*

> Tunnel not reachable: confirm --allow-anonymous flag was used and the agent process is listening on the configured port.

Phase 2 — Register with Agent 365

2.0 — Guardrail: route AI Teammate agents away

This skill is for Agent (Non AI Teammate) registration only. Re-read .a365-workspace-detection.local.json:

• If agentType = "ai-teammate" (stale cache from a prior session, or user picked the wrong skill): abort this skill and tell the user verbatim: *"This agent is registered as an AI Teammate. The make-a365-agent skill handles non-AI-Teammate agents only — appending --authmode obo|s2s to a365 setup all would conflict with --aiteammate. Switch to /agent365:make-ai-teammate instead, which uses a365 setup all --aiteammate --m365 and never passes --authmode."* Do NOT proceed to 2.1.
• If agentType = "system-agent" (or unset): proceed to 2.1.

2.1 — Dry-run preview (REQUIRED before applying anything)

cd "<project_dir>" && a365 setup all --agent-name <agent_name> --dry-run

> --authmode flag: If the user's auth mode is known from .a365-workspace-detection.local.json, append --authmode obo or --authmode s2s to all setup all commands. This controls how the agent identity SP receives permissions (OBO = principal-scoped delegated grants, S2S = application app-role assignments requiring GA).

Show the full dry-run output to the user, then ask:

> "Here's what a365 setup all will create. Does this look correct? Type yes to proceed or no to abort."

• no: Stop. Tell the user "Setup cancelled. Run the make-a365-agent skill again when ready."
• yes: Proceed to 2.2.

2.2 — Apply setup

> ⚠️ a365 setup all is long-running and block-buffers under chat-tool execution. If output stalls, see AGENTS.md § CLI output buffering under chat-tool execution — preferred remediation is run_in_background: true (Claude Code Bash tool); fallback is hand-off to a separate terminal.

Choose the right flags based on the detected agent type:

# Agent (Non AI Teammate) — default
cd "<project_dir>" && a365 setup all --agent-name <agent_name>

# With explicit auth mode (append based on .a365-workspace-detection.local.json authMode)
cd "<project_dir>" && a365 setup all --agent-name <agent_name> --authmode obo
cd "<project_dir>" && a365 setup all --agent-name <agent_name> --authmode s2s

# Custom Engine Agent (CEA) with Teams/Copilot integration — add --m365
cd "<project_dir>" && a365 setup all --agent-name <agent_name> --m365

For CEA agents (usesTeamsOrCopilot = 1): after setup all, also run the bot permission step:

a365 setup permissions bot

This is required for Messaging Bot API grants and must follow setup all (which handles permissions mcp).

This command:

• Creates the Agent 365 Blueprint in Entra ID (agent identity + app registration)
• Grants required Entra ID permissions
• For registration: makes the agent findable in the Agent 365 catalog
• For Custom Engine Agents (--m365): registers the endpoint via MCP Platform

Monitor output carefully:

• The CLI logs progress in numbered steps (e.g. [1/5]). Watch for errors or warnings.
• Existing resources from a previous run are skipped — this is expected behavior.
• If the command fails or prints any error/warning block: show the full CLI output verbatim before consulting the error table below.

> Windows Account Manager (WAM): If you see "Authenticating via Windows Account Manager..." in the output, a native Windows sign-in dialog has appeared. Do NOT kill the process. Tell the user: "A Windows sign-in dialog has appeared — please complete it. Setup will continue automatically after you sign in."

> - If no dialog appears on a headless machine: Ctrl+C, run az login --allow-no-subscriptions to populate the token cache, then retry.

> - Conditional Access Policy (CAP): If WAM/browser auth is blocked by CAP (AADSTS53003, AADSTS53000), the CLI automatically falls back to device code flow — no user action needed.

Handle these conditions:

| Condition | Action |

|-----------|--------|

| Graph API Forbidden / Authorization_RequestDenied | Stop. Resolve permission issue (return to a365-setup Step 2 or grant the role). Then re-run. |

| Interactive browser auth required | If headless, instruct user to use az login --device-code first. |

| managerApplications error / blueprint rejected | Blueprint was created before May 2025 and lacks managerApplications. Delete and re-run a365 setup all, or patch via Graph API. |

| AADSTS700016 / Authorization_IdentityNotFound immediately after blueprint creation | Entra replication lag — the CLI retries automatically with exponential back-off (up to 5× for identity, 12× for blueprint token, 60-second cap). No manual retry needed; wait for the CLI to complete. |

a365 setup all is idempotent — safe to re-run after fixing an issue.

2.3 — Show setup output

After a365 setup all completes, show the user:

• The Setup Summary table from CLI output — verbatim.
• Agent365.Observability.OtelWrite is automatically granted to the agent identity by a365 setup all — no GA consent step required for newly provisioned agents. If the CLI output includes a "Permission Grants" action item (upgrade scenario for pre-1.1 agents), display the PowerShell script verbatim so the user can hand it to a Global Admin.
• Skip the client secret action item entirely. Do not show or mention it.

Mark Todo 1 as completed.

Phase 3 — Add Observability (Optional)

Mark Todo 2 in-progress.

Ask the user:

Your agent is provisioned. Observability lets you track every message, LLM call, and tool
invocation in the Agent 365 portal and Microsoft Defender.

  Would you like to add observability now?
    • yes  — I'll run the instrument-observability skill now
    • skip — you can add it later by running the instrument-observability skill

If yes: Read ${CLAUDE_PLUGIN_ROOT}/skills/instrument-observability/SKILL.md and follow it.

The skill will detect the project language and wire up OTel + A365 tracing exporter.

If skip: Note that the user can run the instrument-observability skill at any time.

Mark Todo 2 as completed when done (or skipped by user).

Phase 4 — Add WorkIQ Tools (Optional)

Mark Todo 3 in-progress.

If authMode = s2s: Skip this phase entirely — WorkIQ is not available for S2S agents (requires a user token). Mark Todo 3 as completed and proceed to Phase 5.

Ask the user:

Would you like to add WorkIQ tools? These give your agent access to Microsoft 365 data —
email, calendar, Teams messages, SharePoint files, OneDrive, and more.

Note: WorkIQ MCP calls use OAuth On-Behalf-Of (OBO) tokens. Users will be prompted to
consent the first time the agent accesses their data.

  • yes  — I'll run the add-workiq-tools skill now
  • skip — you can add it later by running the add-workiq-tools skill

If yes: Read ${CLAUDE_PLUGIN_ROOT}/skills/add-workiq-tools/SKILL.md and follow it.

If skip: Note that the user can run the add-workiq-tools skill at any time.

Mark Todo 3 as completed when done (or skipped by user).

Phase 5 — Final Summary

Show the user a summary:

✅ Agent provisioned with Agent 365!

Your agent now has:
  • Blueprint:       Created in Entra ID (Blueprint ID in `a365.generated.config.json`)
  • Register: Agent appears in the Agent 365 catalog
  [• Observability:  OpenTelemetry + A365 tracing exporter wired]  (if added)
  [• WorkIQ tools:   M365 data access via MCP]                     (if added)

Next steps:
  1. If admin consent was required, ensure a Global Admin has run the PowerShell script.
  2. Test discovery: search for your agent in Microsoft 365 apps.
  3. Add observability:  run the instrument-observability skill  (if not done)
  4. Add WorkIQ tools:   run the add-workiq-tools skill          (if not done)

Error Handling

CLI error surfacing: When any CLI command exits non-zero or prints a warning or error line, show the complete output verbatim in a fenced code block before suggesting a fix. Do not abstract, paraphrase, or discard CLI output — the exact error message is always more useful than a summary. If the error is not in the table below, display it and ask the user how to proceed.

• Run failing commands with -v / --verbose for detailed logs.
• Manage and locate CLI diagnostic logs via a365 logs --help. Log files live at Windows %APPDATA%/a365/logs/, Linux/Mac ~/.config/a365/logs/.
• Most a365 commands are idempotent — safe to re-run after fixing an issue.
• For a full cleanup of a config-free agent: a365 cleanup --agent-name <name>. For granular cleanup: a365 cleanup blueprint, a365 cleanup azure, or a365 cleanup instance. Use only as a last resort.