Runback
Docs menuAgents

    ↑↓ move ↵ open esc close

    Docs The loop

    Agents

    Know which agents act for you, who owns each one, and which models they run on.

    Agents. The loop starts and ends with who is acting. An agent is the name its runs carry; the registry adds who owns it and how much it matters.

    One agent, as the workspace holds itFrom the shared demo workspace
    Agent
    support-refund-agent
    Owner
    Payments Ops (sample)
    Criticality
    high
    Business process
    Customer refunds
    Models the refund run called. Open the run ›
    ModelProviderModel callsTokens
    GPT-4ogpt-4oOpenAI21,030

    The owner, criticality and business process are what the registry row declares. The models are what the run's model calls recorded; Runback prints only the fields those calls carry.

    How to: give an agent an owner

    1. Open Agents › Coverage

      In the app, open Agents › Coverage. It lists every agent that has sent a run, and which ones have no owner yet.

    2. Open the declare form

      Select Declare or update agents.

    3. Add one line per agent

      Each line is name, owner, criticality, optionally followed by system of record, business process, value per run and basis. The name must match the runName (or run_name) your SDK sends. For example:

      ExampleText
      support-agent, Support Ops, high

    4. Declare

      Select Declare.

      Expected output
      The agent's row in Agents and Agents › Coverage shows its owner and criticality. Agents still without an owner stay flagged.

    Agents in the app

    TabWhat it shows
    AgentsThe agents in this workspace and their latest runs
    ModelsWhich models each agent calls, from the spans From Scale
    PromptsVersioned prompts, and which version a run used
    CostSpend per agent, model and team From Scale

    Declaring an agent

    An agent appears in Agents the first time a run arrives under its name. Its owner, criticality and business process are declared on top, and they are what lets the evidence pack answer "who owns each one". Agents › Coverage lists which agents have them.

    Topology and benchmarks

    PageWhat it showsGood to know
    Runs › Topology From ScaleThe fleet as a graph over a 7-, 30- or 90-day window, built from each run's parent_run_idEmpty until subagent runs carry a parent_run_id. Reads at most 20,000 runs per window.
    Benchmarks From ProThe workspace's error rate, policy compliance, average latency and token efficiency beside the median and p90 of anonymised peers in the same verticalUntil enough organisations share the cohort, the comparison is a published-research figure, labelled Reference (published research), with no percentile. Sharing is set in Settings › Sharing › Peer benchmarks and covers error rates and latency only, never prompts or outputs.

    Next

    What an agent did is in Capture.

    Last updated 2 Nov 2026.