Who Gets the First Question?

One team started with Datadog and Snowflake. What would tell me its shared harness was becoming the first place people ask for help?

Plenty of teams have two agents before they’ve chosen one. Datadog comes with an agent and so does Snowflake. A team whose work takes both systems can find that neither product’s agent sees the job through on its own, so it puts a separate harness across the two.

I work on one of these harnesses, so weigh what follows with that in mind. I’ve watched one team start with Datadog and Snowflake. This weekend they connected Confluence, Jira, Microsoft Teams and ServiceNow to the same harness. When someone there has a question and doesn’t yet know which product holds the answer, where do they go first?

The person doing the join

Consider a made-up example. An operations manager opens the morning report and sees that orders were down yesterday.

Did we take fewer orders, or are some of them missing from the report?

Suppose the order service recorded a successful checkout for each order it accepted. Datadog can show those checkouts and when the order exporter failed. In Snowflake, the manager can find the gap in the order records behind the report. She still needs to know whether the missing records are the orders accepted during the failure.

In separate sessions, somebody has to carry the order IDs from one investigation to the other. They also need to carry a less obvious detail. The service timestamps are in UTC and the report closes its day in New York. Lose that detail and two perfectly good queries can produce a plausible discrepancy all by themselves.

I’d be happy to have a specialist find the exporter failure. What I’d like to give the manager is a place where she can ask about orders and come back with a list her team can reconcile. If she has to work out which agent to open next and what to tell it about the previous conversation, she’s doing a fair bit of the work we set out to delegate.

One question, two routes

  1. 1Did we take fewer orders, or are some missing from the report?

    Agent by agent: Datadog agentFinds an exporter failure
      One shared session: Shared session · DatadogFinds an exporter failure
      • 2Are the missing records the orders accepted during the failure?

        Agent by agent: Snowflake agentFinds a gap in the warehouse

        She pastes in

        • Accepted order IDs
        • Failure window
        • UTC vs New York day
        One shared session: Same session · SnowflakeFinds a gap in the warehouse

        Already in the session

        • Order IDs
        • Failure window
        • Time zone
      • 3Who owns the affected accounts?

        Agent by agent: CRM agentLooks up account owners

        She pastes in

        • Missing-order list
        • How it was found
        One shared session: CRM, newly connectedLooks up account owners

        Already in the session

        • Missing-order list
      The operations manager’s three questions, asked two ways. Agent by agent, she carries what she learned into each new conversation. In one shared session it is already there, and connecting the CRM adds a source rather than another conversation to brief.

      Why not use a vendor agent?

      Snowflake’s CoCo can call external tools through MCP. Datadog’s investigation agent can consult source code and telemetry from other systems. Either vendor has a reasonable claim to the manager’s question. If most of her work already starts in Snowflake, connecting Datadog there might suit her quite well.

      Who else would need an account? Datadog meters its agent by usage and doesn’t charge per user. Making it the hub mostly means granting accounts and managing access. Seat-licensed products work differently. Suppose a third-line support engineer wants to check a ServiceNow incident against a Jira issue. She already has seats in both, so either vendor’s agent costs her little beyond usage. Now make one of them the hub for the whole team. If it’s Atlassian’s Rovo, colleagues who work in ServiceNow need Jira seats before they can ask. If it’s ServiceNow, colleagues who only use Jira need ServiceNow seats. Either agent might suit her perfectly. Her manager is choosing for everyone else.

      The independent harness has a cost too. There may be model and query charges, and the usual source-system licences and permissions still apply. But if a colleague already has access to the harness, her manager can spare them a second product seat bought solely to reach its agent.

      I’d also want to know where the platform team’s authority ends. For a job the harness runs, the team has one place to approve models, configure agents and decide which tools each agent may call. Suppose the harness asks Datadog to investigate an alert. The team can govern whether to make that call and what to send with it. Datadog governs the model and processing inside its investigation. A team that uses both will have decisions to make in both places.

      What would make this aggregation?

      In Ben Thompson’s Aggregation Theory, suppliers have a reason to go where customers already go first. Inside a company, a platform team can connect six systems by administrative decision alone. That’s consolidation, and the savings in licences and administration may justify it. For the analogy to Thompson to hold, the people asking the questions have to play a part in where the next connection goes.

      Go back to the operations manager. Suppose she gets her list of missing orders. Now she wants to tell the account owners, whose names are in the CRM. Where would she ask IT to connect it? I’d expect her to want it in the conversation where she already has the order IDs. Opening the CRM’s agent would mean taking the list over and explaining how she got it.

      Say IT adds that connection. The next person who asks about missing orders can find the account owners in the same session. If that makes the harness a better first stop, more people have a reason to start there, including people who haven’t worked out which system they need. The work they bring gives IT a reason to connect the next system there too. That’s the loop I’d look for.

      Who gets the question in Teams?

      Someone can ask a harness agent a question in Teams. With the right permissions, that agent can also look up what the on-call engineer said in a Teams conversation yesterday.

      Copilot can be an aggregator in its own right. Two people can open Teams, one ask Copilot and the other ask a harness agent. Counting visits to Teams would tell me very little about which agent either person trusts with the next question. I’d expect the answer to vary even on the same desk. An on-call engineer might go straight to Datadog for a known alert and ask the shared harness an hour later why yesterday’s orders look low.

      When the next connection goes on the list, I’d like to know who asked for it and what they were trying to finish. Did someone ask for the CRM because they already had the missing orders in front of them?

      References

      1. Snowflake, “Cortex Code and MCP,” Snowflake documentation, docs.snowflake.com.
      2. Datadog, “Configure Bits AI investigations,” Datadog documentation, docs.datadoghq.com.
      3. Datadog, “Pricing: Bits AI credits,” accessed November 2026, datadoghq.com/pricing.
      4. Atlassian, “How Rovo credits work,” Atlassian Support, support.atlassian.com.
      5. ServiceNow Community, “ServiceNow AI-native licensing in 2026: a practical guide to assists, consumption, and governance,” servicenow.com/community.
      6. Datadog, “Calling Bits AI investigations from MCP, API and CLI,” Datadog documentation, docs.datadoghq.com.
      7. Ben Thompson, “Aggregation Theory,” Stratechery, 21 July 2015, stratechery.com.