Governed AI agents for Microsoft Teams.

Cited answers and audited actions. Permission checks happen before tool execution, not after model planning.

In active development — working with design partners

Talk to us about early access

How a request flows

  1. An employee asks in Teams chat — where they already work.
  2. A connector broker checks their identity, role, and permissions before any call is made.
  3. The tool call runs. The model never holds raw credentials.
  4. High-impact actions — writes, sends, exports, permission changes — pause for an approval card, re-verified at the moment it’s clicked.
  5. An audit entry is appended: who asked, who approved, what ran, what it touched.

The refusal is the feature

Employee: Export all customer billing records to a spreadsheet for me.

Agent: I can’t export bulk customer data — that action requires approval under your data policy. I’ve opened an approval request for the billing process owner with the scope you described.

An approval card appears for the process owner: requested action, scope, requester, policy citation. Approve / Deny.

On approval, the scoped export runs. One audit line is appended: who asked, who approved, what ran, what it touched.

A chatbot that can’t refuse isn’t governed. The refusal, the approval path, and the audit line are the product.

Grounded and cited

Answers come from your company’s own documents, policies, and business systems — cited, not improvised from web knowledge. Knowledge is curated by your admins, not crawled.

Approvals

Approval cards arrive in Teams for the people who own the process. Every approval is re-verified at click time — a stale link fails closed. Sensitive changes enforce separation of duties: the person who builds a workflow isn’t the person who approves it.

Audit and data discipline

Every meaningful action produces evidence in an append-only, hash-chained record. Data classification blocks sensitive data from external model providers by default.

Cost governance

Miranda is model-independent: route premium reasoning one way and high-volume work another, set hard spend caps, and see what each agent actually costs — per agent, per run.

Under the hood

LangGraph for durable, checkpointed, resumable workflows. LangChain for model and retrieval components. LangSmith for tracing and evaluation. Agents are built from versioned, reviewable, evaluated workflow components — improvements ship like software releases, not untracked prompt edits. Each customer runs in an isolated, Shiny-operated environment.

Status

Miranda is in active development — working with design partners. Talk to us about early access.